Drex Max v1.1 Decision Model Results

Drex Max v1.1

Explore Drex Max v1.1: exact decision configuration, published specifications, source-linked benchmarks and API tariffs when available.

Published specifications

Provider: Nace.AI
Version: drex-max-v1.1 · Max v1.1 (hosted)
System type: Hosted decision model
Access: Hosted
License: Proprietary
Inputs: text, image
Decision types: Choice, Yes / no, Score
Published size: Not published
Active parameters: Not published
Context / input limit: 262,144 tokens for the whole request (state, images, all questions and formatting)
Announced: 2026-10-28

Model card · Website

Model notes

Nace's hosted Drex Max v1.1, API id drex-max-v1.1, released 28 October 2026 per Nace's models reference. Nace's Drex Max guide calls drex-max (the original Drex Max), drex-max-v1.1 and drex-max-finance separate models with separate serving capacity; all are pinned ids, and drex-latest points to none of them. Nace does not say drex-max is replaced or retired, so no lifecycle link is recorded and both stay listed. Nace's product page captions its family card 'Drex Max v1.1' above the id drex-max; the docs, which give each model its own id, decide. No weights: Hugging Face nace-ai and GitHub nace-ai/drex-decision-models hold only drex-max-v1 (revision a67fa28), and OpenRouter has no nace-ai/drex-max-v1.1 route. Nace publishes no size for v1.1, so the 27B of drex-max-v1 is not copied. Text and PNG, JPEG or WebP images; up to 512 questions and 8 images per request; the 262,144-token limit covers the whole packed request, unlike drex-max's per-question limit. The product page also shows video demonstrations; the API documents images only. Artificial Analysis lists it as 'Drex Max v1.1' next to 'Drex Max', with 262k context and $0.03 per 1M input tokens. Benchmark Heaven (2 November) lists it as a newer version, not yet measured. Nace's results dataset nace-ai/drex-max-v1.1-decision-index-results names the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1 as the weights it ran; that repository returned 401 on 6 November 2026 and the NaceXDeepInfra author lists no public models, so access stays closed.

Checkpoint revision: Hosted revision not published

← All decision models · Compare this decision system →

Published API tariff

$0.03 per 1M input tokens

Nace.AI · checked 2026-11-06

Nace's pricing and models pages, 6 November 2026: $0.03 per 1M input tokens for drex-max-v1.1, output not billed; image tokens count as input. Nace's product page calls v1.1's $0.03 per 1,000 decisions its 50% launch price and gives no end date, so recheck this tariff.

No output-token charge does not mean free requests. These are provider tariffs, separate from benchmark workload costs.

139 recorded observations. Published results retain their source and setup; no Examenos evaluation was run. Missing results are not zero. How to read these comparisons →

Independent · Artificial Analysis

Drex Max v1.1 · Independent · Artificial Analysis
Evaluation	Result	Evidence
AA Decision Index v1.0	64	Setup & source Four equally weighted evaluations: AA-Omniscience-Decision, HLE-MultiChoice (513 text-only questions), When2Call and WildGuardTest. Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Independent AA results. Omniscience splits its 25% between accuracy and non-hallucination. Displayed integer chart values, not undisclosed precise values. No automated AA import or refresh adapter. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-17 · Revision: Manually checked chart · 2026-10-17; 2026-11-01; 2026-11-02; 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
AA Decisions v1 · AA-Omniscience-Decision accuracy	26%	Setup & source AA Decisions v1 · AA-Omniscience-Decision accuracy Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Source-rounded chart label manually verified, not hidden precise data. Five capability components feed the four-benchmark index. Task cost is average workload cost at source current prices, not a token tariff. Time is median across tasks, not a controlled provider benchmark. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-20 · Revision: Manual rendered chart labels · 2026-10-20; 2026-11-01; 2026-11-02; 2026-11-06 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v1 · AA-Omniscience-Decision non-hallucination	70%	Setup & source AA Decisions v1 · AA-Omniscience-Decision non-hallucination Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Source-rounded chart label manually verified, not hidden precise data. Five capability components feed the four-benchmark index. Task cost is average workload cost at source current prices, not a token tariff. Time is median across tasks, not a controlled provider benchmark. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-20 · Revision: Manual rendered chart labels · 2026-10-20; 2026-11-01; 2026-11-02; 2026-11-06 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v1 · HLE-MultiChoice	17%	Setup & source AA Decisions v1 · HLE-MultiChoice Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Source-rounded chart label manually verified, not hidden precise data. Five capability components feed the four-benchmark index. Task cost is average workload cost at source current prices, not a token tariff. Time is median across tasks, not a controlled provider benchmark. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-20 · Revision: Manual rendered chart labels · 2026-10-20; 2026-11-01; 2026-11-02; 2026-11-06 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v1 · When2Call	95%	Setup & source AA Decisions v1 · When2Call Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Source-rounded chart label manually verified, not hidden precise data. Five capability components feed the four-benchmark index. Task cost is average workload cost at source current prices, not a token tariff. Time is median across tasks, not a controlled provider benchmark. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-20 · Revision: Manual rendered chart labels · 2026-10-20; 2026-11-01; 2026-11-02; 2026-11-06 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v1 · WildGuardTest average F1	95%	Setup & source AA Decisions v1 · WildGuardTest average F1 Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Source-rounded chart label manually verified, not hidden precise data. Five capability components feed the four-benchmark index. Task cost is average workload cost at source current prices, not a token tariff. Time is median across tasks, not a controlled provider benchmark. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-20 · Revision: Manual rendered chart labels · 2026-10-20; 2026-11-01; 2026-11-02; 2026-11-06 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v1 · Median time per task	234 ms	Setup & source AA Decisions v1 · Median time per task Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Source-rounded chart label manually verified, not hidden precise data. Five capability components feed the four-benchmark index. Task cost is average workload cost at source current prices, not a token tariff. Time is median across tasks, not a controlled provider benchmark. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-20 · Revision: Manual rendered chart labels · 2026-10-20; 2026-11-01; 2026-11-02; 2026-11-06 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v1 · Cost per million tasks	23 USD/1M tasks	Setup & source AA Decisions v1 · Cost per million tasks Named Drex Max v1.1 by Nace.AI, listed beside the original Drex Max. AA Decisions v1.0. AA's model page gives 262k context and $0.03 per 1M input tokens, matching Nace's drex-max-v1.1. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 6 November 2026; no automated AA adapter. Source-rounded chart label manually verified, not hidden precise data. Five capability components feed the four-benchmark index. Task cost is average workload cost at source current prices, not a token tariff. Time is median across tasks, not a controlled provider benchmark. Superseded on AA by Decision Index v2, first seen 8 November 2026 (aa-decision-index-v2 and aa-decision-v2-*); these v1.0 rows stay as the record of that edition and are no longer refreshed. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-20 · Revision: Manual rendered chart labels · 2026-10-20; 2026-11-01; 2026-11-02; 2026-11-06 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decision Index v2	73	Setup & source Six evaluations: AA-Paperclip 30%, EvalSafe-O*NET-AA 15%, When2Call 10%, AA-Omniscience-Decision v1.1 15%, HLE-MultiChoice v1.1 15% and AA-Symmetry 15%. Each is scored with a Decision Score (accuracy less the Brier score of the chosen answer's probability, 0 to 100), except AA-Symmetry, which uses its Symmetry Score. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published observations; result reuse terms not specified Result source · Benchmark source
AA Decisions v2 · AA-Paperclip Decision Score	73	Setup & source AA Decisions v2 · AA-Paperclip: Decision Score; each document's 15 answers are scored as one decision. 0 to 100, where 100 is certain and right. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · EvalSafe-O*NET-AA Decision Score	88	Setup & source AA Decisions v2 · EvalSafe-O*NET-AA: Decision Score. 0 to 100, where 100 is certain and right. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · When2Call Decision Score	95	Setup & source AA Decisions v2 · When2Call: Decision Score. 0 to 100, where 100 is certain and right. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Omniscience-Decision v1.1 Decision Score	60	Setup & source AA Decisions v2 · AA-Omniscience-Decision v1.1: Decision Score. 0 to 100, where 100 is certain and right. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · HLE-MultiChoice v1.1 Decision Score	47	Setup & source AA Decisions v2 · HLE-MultiChoice v1.1: Decision Score. 0 to 100, where 100 is certain and right. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Symmetry Symmetry Score	85	Setup & source AA Decisions v2 · AA-Symmetry: Symmetry Score. Per question, 0 if the most probable option changes between forms of the question, otherwise 1 less how far that option's probability above chance moves; shown 0 to 100. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · Headline metrics (weighted)	58	Setup & source AA Decisions v2 · Weighted average of the six evaluations' headline metrics, with the index weights. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Paperclip all-pass rate	63%	Setup & source AA Decisions v2 · Share of documents with every decision and extraction right. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · EvalSafe-O*NET-AA accuracy	83%	Setup & source AA Decisions v2 · Share of questions where the answer is the reference consensus's, averaged over the three question types. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · When2Call accuracy	95%	Setup & source AA Decisions v2 · Share of tasks answered with the correct reply. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Omniscience-Decision v1.1 accuracy	28%	Setup & source AA Decisions v2 · Share of questions answered correctly. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Omniscience-Decision v1.1 non-hallucination rate	65%	Setup & source AA Decisions v2 · Share of questions not answered correctly where the model chose a way to say it does not know, or gave no answer. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · HLE-MultiChoice v1.1 accuracy	8%	Setup & source AA Decisions v2 · Share of questions answered with the correct letter. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · HLE-MultiChoice v1.1 non-hallucination rate	50%	Setup & source AA Decisions v2 · Share of questions not answered correctly where the model chose the option that declines, or gave no answer. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Symmetry all-pass permutation rate	55%	Setup & source AA Decisions v2 · Share of questions whose most probable option and probabilities hold across every form. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · Brier score (index weights)	0.16 Brier	Setup & source AA Decisions v2 · Mean squared gap between the chosen answer's probability and whether it is right, 0 to 1: AA-Paperclip, EvalSafe-O*NET-AA, When2Call, AA-Omniscience-Decision v1.1 and HLE-MultiChoice v1.1 averaged with the index weights (AA-Symmetry has no Brier score). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Paperclip Brier score	0.18 Brier	Setup & source AA Decisions v2 · Mean squared gap between the chosen answer's probability and whether it is right, 0 to 1, on AA-Paperclip. Each document's confidence that all 15 answers are right, against whether they are. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · EvalSafe-O*NET-AA Brier score	0.09 Brier	Setup & source AA Decisions v2 · Mean squared gap between the chosen answer's probability and whether it is right, 0 to 1, on EvalSafe-O*NET-AA. Against the reference models' consensus. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · When2Call Brier score	0.04 Brier	Setup & source AA Decisions v2 · Mean squared gap between the chosen answer's probability and whether it is right, 0 to 1, on When2Call. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Omniscience-Decision v1.1 Brier score	0.20 Brier	Setup & source AA Decisions v2 · Mean squared gap between the chosen answer's probability and whether it is right, 0 to 1, on AA-Omniscience-Decision v1.1. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · HLE-MultiChoice v1.1 Brier score	0.23 Brier	Setup & source AA Decisions v2 · Mean squared gap between the chosen answer's probability and whether it is right, 0 to 1, on HLE-MultiChoice v1.1. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · Time per task (index weights)	275 ms	Setup & source AA Decisions v2 · Each evaluation's median time per task, averaged with the index weights. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Time is median across tasks, not a controlled provider benchmark. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Paperclip median time per task	335 ms	Setup & source AA Decisions v2 · Median time per task on AA-Paperclip. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Time is median across tasks, not a controlled provider benchmark. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · EvalSafe-O*NET-AA median time per task	265 ms	Setup & source AA Decisions v2 · Median time per task on EvalSafe-O*NET-AA. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Time is median across tasks, not a controlled provider benchmark. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · When2Call median time per task	284 ms	Setup & source AA Decisions v2 · Median time per task on When2Call. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Time is median across tasks, not a controlled provider benchmark. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Omniscience-Decision v1.1 median time per task	230 ms	Setup & source AA Decisions v2 · Median time per task on AA-Omniscience-Decision v1.1. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Time is median across tasks, not a controlled provider benchmark. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · HLE-MultiChoice v1.1 median time per task	237 ms	Setup & source AA Decisions v2 · Median time per task on HLE-MultiChoice v1.1. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Time is median across tasks, not a controlled provider benchmark. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Symmetry median time per task	241 ms	Setup & source AA Decisions v2 · Median time per task on AA-Symmetry. Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Time is median across tasks, not a controlled provider benchmark. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · Cost per million tasks	102 USD/1M tasks	Setup & source AA Decisions v2 · Average cost per 1 million tasks across all index evaluations at current prices (USD). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. AA prints "*" after this value in every cost chart; no footnote text is rendered. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Task cost is average workload cost at source current prices, not a token tariff. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Paperclip cost per million tasks	282 USD/1M tasks	Setup & source AA Decisions v2 · Average cost per 1 million tasks on AA-Paperclip at current prices (USD). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. AA prints "*" after this value in every cost chart; no footnote text is rendered. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Task cost is average workload cost at source current prices, not a token tariff. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · EvalSafe-O*NET-AA cost per million tasks	55 USD/1M tasks	Setup & source AA Decisions v2 · Average cost per 1 million tasks on EvalSafe-ONET-AA at current prices (USD). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. AA prints "" after this value in every cost chart; no footnote text is rendered. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Task cost is average workload cost at source current prices, not a token tariff. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · When2Call cost per million tasks	32 USD/1M tasks	Setup & source AA Decisions v2 · Average cost per 1 million tasks on When2Call at current prices (USD). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. AA prints "*" after this value in every cost chart; no footnote text is rendered. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Task cost is average workload cost at source current prices, not a token tariff. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Omniscience-Decision v1.1 cost per million tasks	6.40 USD/1M tasks	Setup & source AA Decisions v2 · Average cost per 1 million tasks on AA-Omniscience-Decision v1.1 at current prices (USD). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. AA prints "*" after this value in every cost chart; no footnote text is rendered. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Task cost is average workload cost at source current prices, not a token tariff. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · HLE-MultiChoice v1.1 cost per million tasks	16 USD/1M tasks	Setup & source AA Decisions v2 · Average cost per 1 million tasks on HLE-MultiChoice v1.1 at current prices (USD). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. AA prints "*" after this value in every cost chart; no footnote text is rendered. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Task cost is average workload cost at source current prices, not a token tariff. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source
AA Decisions v2 · AA-Symmetry cost per million tasks	19 USD/1M tasks	Setup & source AA Decisions v2 · Average cost per 1 million tasks on AA-Symmetry at current prices (USD). Named Drex Max v1.1 by Nace.AI. AA Decision Index v2. Exact execution dates, precision and serving details are unpublished. Manual browser observation of a source-rounded chart label on 8 November 2026; no automated AA adapter. AA prints "*" after this value in every cost chart; no footnote text is rendered. Independent AA results from Decision Index v2, first seen on 8 November 2026 (AA's changelog has no entry for it). Displayed chart labels, read by hand, not undisclosed precise values. No automated AA import or refresh adapter. v2 is a new edition with new evaluations and a new scoring rule: never compared with or blended into the v1.0 datasets. Task cost is average workload cost at source current prices, not a token tariff. Checked / reported: 2026-11-08 · Samples: Not published Snapshot: 2026-11-08 · Revision: Manual rendered chart labels · 2026-11-08 · Published summary observations; upstream task/result terms retained Result source · Benchmark source

Official / vendor · Nace.AI vendor product

Peer vendor · Nace.AI vendor release

Drex Max v1.1 · Peer vendor · Nace.AI vendor release
Evaluation	Result	Evidence
Nace Drex for Finance launch · Drex Finance Index	65.80	Setup & source Nace Drex for Finance launch · Drex Finance Index 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Accuracy (3,361 decisions)	77.70%	Setup & source Nace Drex for Finance launch · Accuracy (3,361 decisions) 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Sentiment & tone	69.30	Setup & source Nace Drex for Finance launch · Sentiment & tone 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · News & headlines	71.40	Setup & source Nace Drex for Finance launch · News & headlines 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Banking & support intent	74.50	Setup & source Nace Drex for Finance launch · Banking & support intent 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Documents & extraction	87.80	Setup & source Nace Drex for Finance launch · Documents & extraction 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Financial QA & filings	89.70	Setup & source Nace Drex for Finance launch · Financial QA & filings 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Credit & risk	0.00	Setup & source Nace Drex for Finance launch · Credit & risk 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Published as 0.0: Nace says this is the floor for a model at or below a coin flip, not missing data. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Monetary policy	30.00	Setup & source Nace Drex for Finance launch · Monetary policy 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · ESG & climate	59.40	Setup & source Nace Drex for Finance launch · ESG & climate 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex for Finance launch · Markets & reasoning	43.00	Setup & source Nace Drex for Finance launch · Markets & reasoning 'Drex Max v1.1 Nace' column of the comparison table in Nace's 2 November 2026 Drex for Finance release post; Nace-run. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Nace.AI's 2 November 2026 Drex for Finance release table, checked 6 November 2026. The Drex Finance Index combines 48 public finance test sets held out from training (3,361 decisions); each set is scored as skill above chance, 0 a coin flip and 100 every answer right, with below-chance sets counted as 0. Nace has not published the test-set list, revisions, execution dates or harness. Vendor run; not ranked and not blended with community or independent results. Category score: chance-corrected skill on that category's test sets. Peer columns mapped only where the label names a registered release: Jev 1.14, pplx-decider v1.2, Drex Max v1.1 and Drex Max v1. The 'GPT-6 Luna' and 'GLiDE' columns are not stored: GPT-6 Luna: a bare 'GPT-6 Luna' label names OpenAI's chat model, which is not a Decisions system; Nace writes 'GPT-6 Luna Decisions' when it means that endpoint; GLiDE: no release named, and GLiDE (September 2026) and GLiDE v1.2 are both registered. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-11-02 · Revision: 2026-11-06 · Published observations; result reuse terms not specified Result source · Benchmark source

Official / vendor · Nace.AI results dataset (vendor measured)

Drex Max v1.1 · Official / vendor · Nace.AI results dataset (vendor measured)
Evaluation	Result	Evidence
Nace Drex Max v1.1 results · Knowledge & Reasoning (public skill)	58.7 points	Setup & source Knowledge & Reasoning area skill × 100 (11 benchmarks) from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1. Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as skill 0.5867 on a 0–1 scale; raw (not chance-corrected) 0.6651; coverage 1.0; 11 benchmarks. The README prints 58.7. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Language Understanding (public skill)	74.4 points	Setup & source Language Understanding area skill × 100 (9 benchmarks) from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1. Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as skill 0.7437 on a 0–1 scale; raw (not chance-corrected) 0.8189; coverage 1.0; 9 benchmarks. The README prints 74.4. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Retrieval & Classification (public skill)	71.0 points	Setup & source Retrieval & Classification area skill × 100 (6 benchmarks) from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1. Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as skill 0.7104 on a 0–1 scale; raw (not chance-corrected) 0.7802; coverage 1.0; 6 benchmarks. The README prints 71.0. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Tools & Automation (public skill)	88.0 points	Setup & source Tools & Automation area skill × 100 (5 benchmarks) from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1. Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as skill 0.8802 on a 0–1 scale; raw (not chance-corrected) 0.8961; coverage 1.0; 5 benchmarks. The README prints 88.0. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Arts & Human Taste (public skill)	60.8 points	Setup & source Arts & Human Taste area skill × 100 (6 benchmarks) from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1. Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as skill 0.6082 on a 0–1 scale; raw (not chance-corrected) 0.7562; coverage 1.0; 6 benchmarks. The README prints 60.8. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: Not published Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · ToolRet (nDCG@10)	69.75%	Setup & source ToolRet, nDCG@10, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (685 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.6975309230172237 on a 0–1 scale (nDCG@10). Index component: raw 0.6975, chance-corrected skill 0.6507, chance 0.1341; in the public index. Nace's Drex page shows 71.1 for v1.1 on ToolRet without naming the run; that row is kept separately. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Nace's Drex page shows 71.1 for v1.1 on ToolRet without naming the run; that row is kept separately. Checked / reported: 2026-11-06 · Samples: 685 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · API-Bank (accuracy)	88.39%	Setup & source API-Bank, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (508 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8838582677165354 on a 0–1 scale (accuracy). Index component: raw 0.8839, chance-corrected skill 0.8817, chance 0.0189; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 508 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · BANKING77 (macro-F1)	91.10%	Setup & source BANKING77, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (3,080 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9110202695358018 on a 0–1 scale (macro-F1). Index component: raw 0.911, chance-corrected skill 0.9099, chance 0.0127; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 3080 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · CLINC150+OOS (macro-F1)	94.90%	Setup & source CLINC150+OOS, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,500 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9489644166290192 on a 0–1 scale (macro-F1). Index component: raw 0.949, chance-corrected skill 0.9487, chance 0.006; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5500 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · RouterBench (selected quality (quality objective))	79.56%	Setup & source RouterBench, selected quality (quality objective), from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (10,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.795645 on a 0–1 scale (selected quality (quality objective)). Index component: raw 0.7956, chance-corrected skill 0.5208, chance 0.5735; not in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 10000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Home appliance simulator (case exact accuracy)	98.86%	Setup & source Home appliance simulator, case exact accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (88 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9886363636363636 on a 0–1 scale (case exact accuracy). Index component: raw 0.9886, chance-corrected skill 0.9886, chance 0.0; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 88 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · SGD/SGD-X (macro-F1)	64.88%	Setup & source SGD/SGD-X, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (2,500 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.6487726234097377 on a 0–1 scale (macro-F1). Index component: raw 0.6488, chance-corrected skill 0.4156, chance 0.399; not in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 2500 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · ContractNLI (macro-F1)	83.86%	Setup & source ContractNLI, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (123 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8385590124954193 on a 0–1 scale (macro-F1). Index component: raw 0.8386, chance-corrected skill 0.7665, chance 0.3085; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 123 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · ANLI (macro-F1)	76.50%	Setup & source ANLI, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (3,200 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.7649914358869888 on a 0–1 scale (macro-F1). Index component: raw 0.765, chance-corrected skill 0.648, chance 0.3324; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 3200 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · BPoMP (accuracy)	97.34%	Setup & source BPoMP, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9734 on a 0–1 scale (accuracy). Index component: raw 0.9738, chance-corrected skill 0.9477, chance 0.5; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Humicroedit (accuracy)	67.73%	Setup & source Humicroedit, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (2,628 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.6773211567732116 on a 0–1 scale (accuracy). Index component: raw 0.6773, chance-corrected skill 0.3546, chance 0.5; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 2628 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · POP909-CL (accuracy)	77.20%	Setup & source POP909-CL, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (2,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.772 on a 0–1 scale (accuracy). Index component: raw 0.7707, chance-corrected skill 0.7689, chance 0.0078; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 2000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · cfcolor (accuracy)	65.34%	Setup & source cfcolor, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.6534 on a 0–1 scale (accuracy). Index component: raw 0.6784, chance-corrected skill 0.3568, chance 0.5; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · MMLU (accuracy)	87.36%	Setup & source MMLU, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (14,033 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8735836955747167 on a 0–1 scale (accuracy). Index component: raw 0.8736, chance-corrected skill 0.8315, chance 0.25; not in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 14033 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · GPQA Diamond (accuracy)	54.59%	Setup & source GPQA Diamond, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (196 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.5459183673469388 on a 0–1 scale (accuracy). Index component: raw 0.5459, chance-corrected skill 0.3946, chance 0.25; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 196 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · WinoGrande (accuracy)	93.45%	Setup & source WinoGrande, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (1,267 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9344909234411997 on a 0–1 scale (accuracy). Index component: raw 0.9345, chance-corrected skill 0.869, chance 0.5; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 1267 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · HellaSwag (accuracy)	97.05%	Setup & source HellaSwag, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (10,042 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9705238000398327 on a 0–1 scale (accuracy). Index component: raw 0.9705, chance-corrected skill 0.9607, chance 0.25; in the public index. Nace's Drex page shows 98.1 for v1.1 on HellaSwag without naming the run; that row is kept separately. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Nace's Drex page shows 98.1 for v1.1 on HellaSwag without naming the run; that row is kept separately. Checked / reported: 2026-11-06 · Samples: 10042 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · ChessBench (accuracy)	25.80%	Setup & source ChessBench, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.258 on a 0–1 scale (accuracy). Index component: raw 0.258, chance-corrected skill 0.1918, chance 0.0819; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · MuSR (accuracy)	71.54%	Setup & source MuSR, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (752 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.7154255319148937 on a 0–1 scale (accuracy). Index component: raw 0.7154, chance-corrected skill 0.5475, chance 0.371; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 752 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · SATA-Bench (case exact accuracy)	38.06%	Setup & source SATA-Bench, case exact accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (1,650 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.3806060606060606 on a 0–1 scale (case exact accuracy). Index component: raw 0.3806, chance-corrected skill 0.3724, chance 0.0131; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 1650 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · SimpleBench (accuracy)	30.00%	Setup & source SimpleBench, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (10 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.3 on a 0–1 scale (accuracy). Index component: raw 0.3, chance-corrected skill 0.16, chance 0.1667; not in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 10 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · BRIGHT (nDCG@10)	49.61%	Setup & source BRIGHT, nDCG@10, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (220 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.4960763085785154 on a 0–1 scale (nDCG@10). Index component: raw 0.4961, chance-corrected skill 0.4299, chance 0.116; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 220 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Amazon ESCI (macro-F1)	59.96%	Setup & source Amazon ESCI, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.5996309606305666 on a 0–1 scale (macro-F1). Index component: raw 0.5996, chance-corrected skill 0.4978, chance 0.2027; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · ACOS (per-review F1)	53.51%	Setup & source ACOS, per-review F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (1,565 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.5350859022513434 on a 0–1 scale (per-review F1). Index component: raw 0.5351, chance-corrected skill 0.5202, chance 0.031; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 1565 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · FinEntity (macro-F1)	94.75%	Setup & source FinEntity, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (979 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9474797135055434 on a 0–1 scale (macro-F1). Index component: raw 0.9475, chance-corrected skill 0.9228, chance 0.3201; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 979 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · VAST (macro-F1)	81.02%	Setup & source VAST, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (3,006 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8101802992851731 on a 0–1 scale (macro-F1). Index component: raw 0.8102, chance-corrected skill 0.7153, chance 0.3333; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 3006 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · NLI4CT (macro-F1)	86.54%	Setup & source NLI4CT, macro-F1, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,500 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8653669254294748 on a 0–1 scale (macro-F1). Index component: raw 0.8654, chance-corrected skill 0.7381, chance 0.486; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5500 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · CLadder (accuracy)	92.78%	Setup & source CLadder, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9278 on a 0–1 scale (accuracy). Index component: raw 0.9278, chance-corrected skill 0.8556, chance 0.5; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · HLE (accuracy)	15.57%	Setup & source HLE, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (501 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.15568862275449102 on a 0–1 scale (accuracy). Index component: raw 0.1557, chance-corrected skill 0.0, chance 0.1641; in the public index. Nace's Drex page shows 18.2 for v1.1 on HLE without naming the run; that row is kept separately. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Nace's Drex page shows 18.2 for v1.1 on HLE without naming the run; that row is kept separately. Checked / reported: 2026-11-06 · Samples: 501 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Habermas Machine (accuracy)	62.83%	Setup & source Habermas Machine, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (1,676 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.6282816229116945 on a 0–1 scale (accuracy). Index component: raw 0.6283, chance-corrected skill 0.4605, chance 0.311; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 1676 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · GSM8K (accuracy)	83.89%	Setup & source GSM8K, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (2,638 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8388931008339652 on a 0–1 scale (accuracy). Index component: raw 0.8389, chance-corrected skill 0.8068, chance 0.25; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 2638 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · PhishNChips phishing decisions (accuracy)	84.20%	Setup & source PhishNChips phishing decisions, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (2,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.842 on a 0–1 scale (accuracy). Index component: raw 0.842, chance-corrected skill 0.684, chance 0.5; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 2000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · BBH fixed-option tasks (accuracy)	93.83%	Setup & source BBH fixed-option tasks, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (5,507 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.9382603958598148 on a 0–1 scale (accuracy). Index component: raw 0.9383, chance-corrected skill 0.9106, chance 0.3101; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 5507 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · HoVer claim verification (accuracy)	88.05%	Setup & source HoVer claim verification, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (4,000 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8805 on a 0–1 scale (accuracy). Index component: raw 0.8805, chance-corrected skill 0.761, chance 0.5; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 4000 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · New Yorker caption matching (accuracy)	80.87%	Setup & source New Yorker caption matching, accuracy, from Nace's own Decision Index 0.3 public-suite run of Drex Max v1.1 (528 scored requests). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on one NVIDIA B200 per worker; Decision Index kit c254f8d, public suite release-v3, 140,178 scoreable requests, edition 0.3 scoring. Not a run of the hosted API. Published as 0.8087121212121212 on a 0–1 scale (accuracy). Index component: raw 0.8087, chance-corrected skill 0.7609, chance 0.2; in the public index. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Text run: the frozen public suite release-v3 (rows_sha256 b2b56d6fb636…, 140,620 requests, 140,178 scoreable, 43 benchmarks, 37 in the index), the same files as the original Drex Max's results dataset, scored on 27 October 2026 with Decision Index kit c254f8d (--edition 0.3) and labelled the Decision Index 0.4 public suite. The public index is 100 × the weighted mean of five area skills; each benchmark is chance-corrected, (score − chance) / (1 − chance) clipped to 0–1, and coverage-adjusted. scores.json also gives a raw index of 77.92 and a breadth skill of 70.15, not stored as columns. Its latency fields are in-process wall time under batched inference on one NVIDIA B200 per worker, not serving latency, and are not stored. Checked / reported: 2026-11-06 · Samples: 528 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Public vision index	75.78	Setup & source Public vision index from Nace's own run of Drex Max v1.1 on the Decision Index vision suite. Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as 75.78447602638617. Nace's Drex page shows 75.9 for v1.1's vision panel; its accuracies give 75.89 with this run's chance levels and weights, so the panel is a different vision run. Both are kept. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 12210 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · CV-Bench (vision skill)	77.4 points	Setup & source CV-Bench chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (2,038 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.7741591678696222 on a 0–1 scale; accuracy 0.8631010794896958, chance 0.393825646058227, index weight 1.0; the README prints 77.4. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 2038 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · BLINK (vision skill)	61.3 points	Setup & source BLINK chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (961 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.6133709222714459 on a 0–1 scale; accuracy 0.7502601456815817, chance 0.3540582726326743, index weight 1.0; the README prints 61.3. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 961 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · RealWorldQA (vision skill)	73.1 points	Setup & source RealWorldQA chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (589 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.7308670520231213 on a 0–1 scale; accuracy 0.8353140916808149, chance 0.38808715336728916, index weight 1.0; the README prints 73.1. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 589 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · CharXiv (vision skill)	80.0 points	Setup & source CharXiv chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (532 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.7994987468671679 on a 0–1 scale; accuracy 0.849624060150376, chance 0.25, index weight 1.0; the README prints 79.9. Nace's Drex page shows accuracy 86.6% for v1.1; this run's accuracy is 84.96%. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 532 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · InfographicVQA (vision skill)	98.4 points	Setup & source InfographicVQA chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (244 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.9836065573770493 on a 0–1 scale; accuracy 0.9877049180327869, chance 0.25, index weight 1.0; the README prints 98.4. Nace's Drex page shows accuracy 98.4% for v1.1; this run's accuracy is 98.77%. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 244 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Mind2Web (vision skill)	86.3 points	Setup & source Mind2Web chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (973 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.8629667694415896 on a 0–1 scale; accuracy 0.8972250770811921, chance 0.25, index weight 1.0; the README prints 86.3. Nace's Drex page shows accuracy 89.2% for v1.1; this run's accuracy is 89.72%. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 973 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Winoground (vision skill)	86.8 points	Setup & source Winoground chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (800 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.8674999999999999 on a 0–1 scale; accuracy 0.93375, chance 0.5, index weight 1.0; the README prints 86.8. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 800 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · KIE (CORD+FUNSD) (vision skill)	93.2 points	Setup & source KIE (CORD+FUNSD) chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (2,465 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.9318751043579897 on a 0–1 scale; accuracy 0.9586206896551724, chance 0.39259634888438133, index weight 1.0; the README prints 93.2. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 2465 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · Moderation (Hateful Memes) (vision skill)	45.6 points	Setup & source Moderation (Hateful Memes) chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (1,213 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.4558944765045343 on a 0–1 scale; accuracy 0.7279472382522671, chance 0.5, index weight 1.0; the README prints 45.6. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 1213 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · R-Bench-M (vision skill)	50.4 points	Setup & source R-Bench-M chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (665 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.5036847646262596 on a 0–1 scale; accuracy 0.5864661654135338, chance 0.16679197994987469, index weight 0.5; the README prints 50.4. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 665 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source
Nace Drex Max v1.1 results · MMMU-Pro vision (vision skill)	47.0 points	Setup & source MMMU-Pro vision chance-corrected skill × 100 from Nace's own run of Drex Max v1.1 on the Decision Index vision suite (1,730 rows). Nace-run Drex Max v1.1 (engine drex-max-v1.1), weights loaded locally from NaceXDeepInfra/Drex-Max-v1.1 (not public) on NVIDIA B200; Decision Index kit c254f8d, public vision suite, 12,210 rows, images capped at 1,638,400 pixels. Not a run of the hosted API. Published as skill 0.4696209341880083 on a 0–1 scale; accuracy 0.5375722543352601, chance 0.12811840535828975, index weight 0.25; the README prints 47.0. Nace.AI's own Decision Index run of Drex Max v1.1, published in the Hugging Face dataset nace-ai/drex-max-v1.1-decision-index-results (revision 1c6b41a, 27 October 2026; read 6 November 2026). The run names engine drex-max-v1.1 and loaded weights locally from the Hugging Face repository NaceXDeepInfra/Drex-Max-v1.1, which is not public; no weights hash is published, so the run is tied to the hosted release by Nace's naming only. Values that Nace's Drex page already shows for this run (public index 70.8, MMLU-Pro, CRUXEval, ARC-Challenge, ARC-Easy, BFCL, When2Call and RAGTruth) stay on the product rows and are not stored twice. Vendor run; not ranked and not blended with community or independent results. Vision run: the public Decision Index vision suite (rows_sha256 f5f1cfb5cd92…, 12,210 rows, identical to the original Drex Max's vision run), images at native resolution capped at 1,638,400 pixels, scored with the kit's vision score (edition 0.3, Nace's '0.4 Vision board'). The column is chance-corrected skill × 100; R-Bench-M weighs 0.5 and MMMU-Pro vision 0.25 in the index, the rest 1. The run was replaced on 27 October: the first upload (commit bc74e68) gave 75.39 on a locally rebuilt suite whose Mind2Web images and 52 InfographicVQA images differed from the reference build; the stored run passes vision verify. Nace's Drex page shows accuracy, not skill, and its vision panel is a different run (see each row). Checked / reported: 2026-11-06 · Samples: 1730 Snapshot: 2026-10-27 · Revision: 1c6b41a03b66d13f351b611c3193a0893b98bf11 · Published observations; result reuse terms not specified Result source · Benchmark source