ExamenosThe benchmark library
Choose what to compare.
Browse by category, then open a benchmark to see every model’s results.
Data coverage
Choose a model to see which benchmarks have recorded results and which are missing.
Explore coverage → Explore changes by dateResults over time
Choose a benchmark, then drag its timeline to see the scores we had recorded at each date.
Open benchmark timelines →Source-linked benchmark results
GPQA Diamond
Graduate-level science QA
52 / 100 models with results →Official / vendorDeepSWE v1.1
Long-horizon agentic software engineering on original tasks from active open-source repos (Datacurve DeepSWE).
48 / 100 models with results →Official / vendorTerminal-Bench 2.1
Earlier Terminal-Bench agent suite (shorter/earlier protocol than 3.0/4.0); keep separate from later versions.
48 / 100 models with results →Official / vendorTerminal-Bench 4.0
Multi-hour agentic terminal/coding tasks: shell, tools, and long-running project work under a hard time budget.
38 / 100 models with results →Official / vendorMMMU Pro (no tools)
MMMU-Pro without tools.
33 / 100 models with results →Official / vendorτ³-bench Banking
Tau-3 Banking agentic evaluation (AA / vendor-cited).
33 / 100 models with results →Official / vendorAutomationBench v1.0.6
General agent automation across multi-step workflows (forms, apps, scripts) rather than pure coding.
30 / 100 models with results →Official / vendorAA Intelligence Index v4.3.2 (vendor-cited)
Artificial Analysis Intelligence Index v4.3.2 exactly as cited by a model vendor. AA rescales and reweights each index version, so values are not comparable with other versions. Independent AA observations stay separate.
29 / 100 models with results →Official / vendorSWE-bench Pro (revision unspecified)
Long-horizon software engineering on real GitHub issues beyond the classic SWE-bench Verified set.
29 / 100 models with results →Official / vendorFrontierCode 2.0
Cognition's 132-task mergeability benchmark, released 27 October 2026: maintainer-written, deliberately underspecified prompts, more repositories, stricter blocking criteria and larger task scope. Weighted rubric score; runs that fail a blocking criterion score 0. Each model runs in its own named harness. Not comparable with FrontierCode 1.1 or Artificial Analysis' FrontierCode runs.
28 / 100 models with results →Official / vendorSWE-bench Verified
Human-validated 500-instance subset of SWE-bench for reliable coding-agent evaluation (distinct from SWE-bench Pro).
28 / 100 models with results →IndependentIntelligence Index v4.3.2 · Artificial Analysis
Artificial Analysis Intelligence Index v4.3.2. Every stored observation was captured while v4.3.2 was current (19 September to 27 October 2026); v5.0 and later versions are stored as separate metrics because AA rescales and reweights each version. Exact execution dates and serving details may be unpublished. This is not a controlled time comparison.
40 / 100 models with results →IndependentVals Index (legacy, before v2.2) · Vals AI
Legacy Vals Index observations recorded before the v2.2 benchmark revision. Preserve these historical results separately: v2.2 changes components, test populations and grading. This is not a controlled before/after comparison with v2.2.
27 / 100 models with results →IndependentVals Index v2.2 · Vals AI
GDP-weighted aggregate with revised Finance v2.1, EMB v1.1, HLAB v1.1, Code Migration v1.1 and all-pass Tax. Old Index versions are not comparable. Source default includes provider-refusal fallback tasks for Opus/Sonnet; failure-controlled scores are separate.
35 / 100 models with results →IndependentBug Hunt Bench · Bugs fixed (105 total) · Bug Hunt Bench
Recorded Bugs fixed /105 measurements from Bug Hunt Bench, with source links, evaluation dates and available effort or harness records. Vendor results remain separate.
30 / 100 models with results →IndependentARC-AGI-2 · ARC Prize
ARC-AGI-2: harder successor to ARC-AGI-1 with more compositional, less brute-forceable grid tasks. Still few-shot rule induction on novel puzzles; designed to stay unsaturated as models saturate v1. ARC Prize reports Semi-Private verified %.
23 / 100 models with results →The interactive library loads every tracked metric. Browse public benchmark tables or model profiles.
| Provider | Published / checked | Evidence |
|---|