Sage-A1.5 Benchmarks, Specifications & Availability

Sage-A1.5

Explore Sage-A1.5 from Shanghai AI Laboratory: published specifications, source-linked vendor benchmarks, independent evaluator coverage and recorded pricing when available.

Published specifications

Provider: Shanghai AI Laboratory
Access: Open weights
License: Apache-2.0
Context window: 256K
Total parameters: 35B
Active parameters: 3B
Released: 2026-10-15
Modalities: text, image
Family: Sage

Hugging Face · Announcement · Website

Model notes

Open BF16 weights verified 21 October. Model card/report publish 35B MoE and 3B active (Figure 9); HF exact 35,951,822,704 total. Configuration 262144 tokens; no measured full-context guarantee. Initial weights 15 October 2026, report 20 October, card updates 21 October; announced_date uses initial public weights fallback because distinct dated launch article unavailable. Agentic Orchestration and Verification are opt-in; teacher/SFT ablations are not this released checkpoint. Supports vision; video architecture present but dedicated video performance not asserted.

Compare Sage-A1.5 with other models →Compare subscription plans Explore data coverage

Pricing · OpenRouter

No OpenRouter pricing is recorded for this model. Missing rates are not free.

Official / vendor benchmarks

Default headline records. Own-vendor, peer-vendor and third-party provenance remain visible in evidence; configurations may differ.

Official / vendor headline scores; expand evidence for every record
Benchmark / evaluator	Headline score	Evidence
GAIA (Sage protocol)	95.1%unknown effort	All 1 recorded result & sources 95.1% · raw 95.1 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Search/visit/Python tools; top 50 search; visit uses dedicated summarization model. Context reset at 300 calls or context limit, up to 5 rounds. GAIA split unpublished. https://arxiv.org/html/2610.28578v1#S3.T1
BrowseComp	81.7%unknown effort	All 2 recorded results & sources 81.7% · raw 81.7 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Search/visit/Python; top 50 search, summarizer; context reset at 300 calls/context limit, up to 5 rounds. Standard mode without agentic orchestration. https://arxiv.org/html/2610.28578v1#S3.T1 82.8% · raw 82.8 % Alternative · unknown effort · Own vendor Source/record date: 2026-10-21 Harness: Agentic Orchestration enabled, spawn_subagent; subagents summarize after 50 turns. Table 4 final model, not teacher. Search/visit/Python; top 50 search, summarizer; context reset at 300 calls/context limit, up to 5 rounds. Standard mode without agentic orchestration. https://arxiv.org/html/2610.28578v1#S3.T4
DeepResearch Bench II (Sage protocol)	43.7%unknown effort	All 2 recorded results & sources 43.7% · raw 43.7 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Expert-rubric report generation; same search/visit/Python and reset protocol; no orchestration. https://arxiv.org/html/2610.28578v1#S3.T1 47.5% · raw 47.5 % Alternative · unknown effort · Own vendor Source/record date: 2026-10-21 Harness: Agentic Orchestration enabled, spawn_subagent; subagents summarize after 50 turns. Table 4 final model, not teacher. Expert-rubric report generation; same search/visit/Python and reset protocol; no orchestration. https://arxiv.org/html/2610.28578v1#S3.T4
XBench-DS-2510 (Sage protocol)	92%unknown effort	All 1 recorded result & sources 92% · raw 92 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. October2025 xbench DeepSearch profession-aligned question set; same search/visit/Python harness. https://arxiv.org/html/2610.28578v1#S3.T1
MLE-Bench Lite (Sage protocol)	65.2%unknown effort	All 1 recorded result & sources 65.2% · raw 65.2 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 22 Kaggle competitions, fraction earning at least bronze;3 seeds, dedicated H200, 12 h/task. https://arxiv.org/html/2610.28578v1#S3.T1
SciCode (subtask)	44.3%unknown effort	All 1 recorded result & sources 44.3% · raw 44.3 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 288 test subproblems; scientific background provided, sequential pass@1; all hidden unit tests must pass. https://arxiv.org/html/2610.28578v1#S3.T1
SWE-bench Pro (revision unspecified)	61.2%unknown effort	All 1 recorded result & sources 61.2% · raw 61.2 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Mini-SWE-Agent; offline, restricted Git history; card temperature 1, top_p 0.95,200K context and automatic compaction; revision unspecified. https://arxiv.org/html/2610.28578v1#S3.T1
SWE-bench Multilingual	72.7%unknown effort	All 1 recorded result & sources 72.7% · raw 72.7 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Mini-SWE-Agent; nine languages; offline/restricted Git; card temperature 1, top_p 0.95,200K context/automatic compaction. https://arxiv.org/html/2610.28578v1#S3.T1
IFBench (prompt strict)	81.6%unknown effort	All 1 recorded result & sources 81.6% · raw 81.6 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 294 prompts; all constraints satisfied, strict prompt-level accuracy. https://arxiv.org/html/2610.28578v1#S3.T1
AA-LCR	81%unknown effort	All 1 recorded result & sources 81% · raw 81 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 100 multi-document questions, reference LLM judge, accuracy averaged 3 runs; source version not explicit. https://arxiv.org/html/2610.28578v1#S3.T1
FrontierScience Olympiad (Sage protocol)	81%unknown effort	All 1 recorded result & sources 81% · raw 81 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Olympiad-level physics/chemistry/biology; accuracy. https://arxiv.org/html/2610.28578v1#S3.T1
FrontierScience Research (Sage protocol)	48.3%unknown effort	All 1 recorded result & sources 48.3% · raw 48.3 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. PhD-level research subtasks; accuracy. https://arxiv.org/html/2610.28578v1#S3.T1
HiPhO (Sage protocol)	45.3 pointsunknown effort	All 1 recorded result & sources 45.3 points · raw 45.3 points Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 13 physics competitions from 2024–2025; mean competition score, not exact-match accuracy. https://arxiv.org/html/2610.28578v1#S3.T1
Humanity's Last Exam (with tools)	55.1%unknown effort	All 1 recorded result & sources 55.1% · raw 55.1 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Academic multimodal questions with tools; exact split/version unreported. https://arxiv.org/html/2610.28578v1#S3.T1
CMT (Sage protocol)	80%unknown effort	All 1 recorded result & sources 80% · raw 80 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Condensed Matter Theory benchmark; accuracy. https://arxiv.org/html/2610.28578v1#S3.T1
CritPt	20%unknown effort	All 1 recorded result & sources 20% · raw 20 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Model-generated solutions submitted to Artificial Analysis for official grading; vendor-reported accuracy; run date unknown. https://arxiv.org/html/2610.28578v1#S3.T1
ALE-CLI strict full score (Sage protocol)	19.1%unknown effort	All 1 recorded result & sources 19.1% · raw 19.1 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 105 tasks / 12 domains, Claude Code; clean CPU sandbox 8 h, max output 64K / thinking 32K. Task passed only at full score. Separate from weighted/partial ALE. https://arxiv.org/html/2610.28578v1#S3.T1
VitaBench (Sage protocol)	45.4%unknown effort	All 1 recorded result & sources 45.4% · raw 45.4 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Vendor Table 1; version/split unreported. https://arxiv.org/html/2610.28578v1#S3.T1
PinchBench v1 (Sage protocol)	90%unknown effort	All 1 recorded result & sources 90% · raw 90 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. PinchBench-v1,23 tasks, OpenClaw;5 runs;262K context, 65K max output, 3–15 min; DeepSeek-V4-Flash judge. https://arxiv.org/html/2610.28578v1#S3.T1
ChemCoTBench (Sage protocol)	66.5%unknown effort	All 1 recorded result & sources 66.5% · raw 66.5 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 1,495 frozen examples / 19 subtasks; original query, no demos/tools;65,536 max output; frozen chemical scorer; unweighted subtask mean. https://arxiv.org/html/2610.28578v1#S3.T1
MatTools (Sage protocol)	42.6%unknown effort	All 1 recorded result & sources 42.6% · raw 42.6 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. 138 subtasks / 49 problems; pymatgen Python in Docker, official verification; task success. https://arxiv.org/html/2610.28578v1#S3.T1
OmniDocBench overall (revision unspecified) (Sage protocol)	91.5%unknown effort	All 1 recorded result & sources 91.5% · raw 91.5 % Headline · unknown effort · Own vendor Source/record date: 2026-10-21 Sage-A1.5 report Table 1. Final released checkpoint. Official three-component overall: text normalized edit similarity, table TEDS, formula CDM; revision not published, never label 1.5/1.6. https://arxiv.org/html/2610.28578v1#S3.T1

Independent evaluators

Evaluator harnesses are distinct from vendor measurements. Missing coverage is not a failed test.

No independent evaluator results are recorded for this model.

Read how we select and source scores or the comparison guide.