Benchmarks
Every benchmark has its own leaderboard and everything its publisher releases beyond the score. Where one measures cost per task, Priors adjusts it for subscription plans, so you see what a model costs to actually use.
318 benchmarksup to 445 models each111 publishers
/
How scores work318 benchmarks
- Terminal-Bench v4-AAAA's 66-task Terminal-Bench v4.0, three repeats, average pass@1. Distinct from the upstream leaderboard and v2.1.
artificialanalysis.aiTop model: Kimi K3.1151 models23 mixesQuality 100
- HLE-DiamondHLE-Diamond: the refined 1,000-question version of Humanity's Last Exam from CAIS and Scale AI, split evenly between reasoning and knowledge questions; accuracy.
- FrontierCode 2.0-AAAA's run of Cognition's FrontierCode 2.0 (132 private tasks in real open-source repositories, three attempts each): would a maintainer merge the change? An attempt failing any blocking check scores 0. Separate from Cognition's own leaderboard, which uses different harnesses.
- FrontierMath Tier 4 (v2)1 sub-boardEpoch AI's FrontierMath Tier 4, version 2: expert-written mathematics problems evaluated with Python tools. Accuracy on the private set, not the public sample questions. Data: Epoch AI, CC BY.
epoch.aiTop model: GPT-6.1 Sol68 models9 mixesQuality 90
- GDP.pdfGDP.pdf: document-heavy professional tasks over PDFs graded against rubric criteria; share of tasks passing all criteria.
- ProgramBench v1 Fully Resolved2 sub-boardsShare of public ProgramBench v1 tasks where the implementation passes every hidden behavioral test.
vals.aiTop model: Opus 5.562 models4 mixesQuality 99
- AutomationBench-AAAutomationBench-AA: building workflow automations; partial-credit score.
- Terminal-Bench-Science-AATerminal-Bench-Science run by AA: 70 expert-curated scientific research workflows across five domains in a terminal; a task passes only if every one of its tests passes. Pass@1 over three repeats.
- AA-Briefcase v1.1AA-Briefcase: agentic office-work tasks over real files (xlsx/pptx/docx/pdf) judged pairwise; reported as Elo.
- DeepSWE v1.1Pass@1 on 113 original long-horizon software engineering tasks across 91 repos and 5 languages, run with mini-swe-agent in isolated environments (v1.1 fixes dependency drift and flaky tasks).
deepswe.datacurve.aiTop model: GPT-6 Astra28 models8 mixesQuality 76
- ARC-AGI-3ARC-AGI-3 semi-private set: interactive game-like environments where the agent must learn rules on the fly; percent of possible score.
arcprize.orgTop model: GPT-6 Astra24 models7 mixesQuality 78
- BullshitBench v2Share of 100 nonsense prompts (software, finance, legal, medical, physics) where the model clearly calls out the nonsense instead of confidently answering; 3-judge panel, mean aggregation. Reported per reasoning level.
petergpt.github.ioTop model: Opus 4.8108 models5 mixesQuality 80
- Terminal-Bench 4.0Resolution rate of agent+model systems on 66 professional computer-work tasks in a terminal (5 trials per task, 8-hour timeout). Rows are harness+model pairs.
tbench.aiTop model: Opus 5.523 models10 mixesQuality 67
- MLCR-AAMedical Long Context Reasoning (MLCR-AA): reasoning over long medical records; overall score.
- Arena WebDev3 sub-boardsHuman preference between agentic front-end web development outputs; Bradley-Terry rating.
arena.aiTop model: Opus 5.5113 models13 mixesQuality 62
- SimpleBench200+ multiple-choice questions on spatio-temporal reasoning, social intelligence and adversarial trick questions where unspecialized humans outperform frontier models; accuracy averaged over 5 runs.
simple-bench.comTop model: Opus 5.596 models9 mixesQuality 66
- Terminal-Bench Science (Vals)Research workflow tasks contributed by practicing scientists across five domains
- Agent Arena6 sub-boardsArena.ai Agent Arena: causal 'net improvement' a model contributes to real-world agentic sessions (task success, praise vs complaint, steerability, bash recovery, tool hallucination), estimated from millions of live Agent Mode sessions. Percent with 95% CI; not an Elo.
- FrontierCode 2.0Mergeability of coding-agent changes on private maintainer-authored tasks, graded by unit tests, rubrics and verifiers (Cognition).
cognition.comTop model: Opus 5.525 models4 mixesQuality 73
- GDPval-AA v2.1GDPval-AA: economically valuable knowledge-work tasks across occupations, graded pairwise by AA; reported as Elo.
- FrontierSWE v2Cognition's ultra-long-horizon software engineering benchmark: 34 tasks graded 0-100 on implementation, performance and research; mean@5 with a 20-hour budget per task in the Proximus harness at max reasoning effort.
frontierswe.comTop model: GPT-6 Astra22 models4 mixesQuality 67
- Terminal-Bench 4.0 (Vals)Vals AI's Terminal-Bench 4.0 aggregate scores from its external-runner leaderboard. Vals publishes a separate system set and score series from the tbench.ai harness-plus-model leaderboard, so the two sources remain distinct.
- WeirdML v3Fully agentic: 11 hand-made ML and data-analysis tasks run in coding-agent harnesses, where models must explore unfamiliar data and build pipelines with limited data, unspecified goals or little feedback. Overall effective score (95% interval).
htihle.github.ioTop model: GPT-6 Astra25 models4 mixesQuality 65
- AA Intelligence Index v5.0Artificial Analysis' composite intelligence score (v5.0), combining 9 evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), Terminal-Bench-Science 0.1.
- Harvey LAB-AA v1.1Harvey LAB-AA v1.1: 120 private agentic legal tasks built by Harvey. Hallucination-Gated All-Pass Rate: a task counts only when a three-judge panel passes every rubric criterion and the deliverables contain no material hallucination. Not comparable with v1.0 scores.
- AA-Omniscience v1.19 sub-boardsAA-Omniscience v1.1 Index (-100..100): knowledge accuracy minus a hallucination penalty. v1.1 drops saturated and public questions, adds new ones and revises ambiguous ones; scores run about 2 points below v1.
- T3 Code usage shareShare of all turns sent in T3 Code (2026-10-04 to 2026-11-02), from T3's own telemetry via Slopalytics. Revealed preference among T3 Code users, not a quality measure: price, plans and harness defaults drive it too. Model-level, no reasoning effort.
slopalytics.comTop model: Opus 5.591 models1 mixQuality 76
- Code2Video-AACode2Video-AA: AA's run of HeyGen's Code2Video Bench. Models turn 114 design briefs into short product and brand videos in HyperFrames HTML; Elo from blind pairwise votes.
- AutomationBench (Zapier)Zapier's agent benchmark (v1.0.6): end-to-end business workflows across 47 real tools in six functions, graded strictly on the final environment state (every assertion must pass). Private held-out set.
zapier.comTop model: Gemini 4 Carbon40 models2 mixesQuality 68
- AA-LCR v1.1AA-LCR: long-context reasoning over ~100k-token document sets; pass rate.
- AA-3DAA-3D: agents in an offline sandbox turn a written brief into an interactive three.js scene; Elo from blind pairwise votes by AA's recruited panel.
- VoxelBench (text)Models build 3D voxel scenes from text prompts; humans vote head-to-head on the renders. Glicko-2 rating.
voxelbench.aiTop model: GPT-6 Astra54 models4 mixesQuality 58
- Chess Puzzles (Epoch AI)Epoch AI's chess puzzles: the model finds the best move in positions with a single correct line. Score is the share of puzzles solved. Data: Epoch AI, CC BY.
- Blueprint-Bench 2Spatial reasoning: reconstruct 2D floor plans from ~20 apartment photos, scored on room connectivity and scale (random baseline normalized to 0).
andonlabs.comTop model: Gemini 4 Carbon30 models1 mixQuality 72
- KernelBench CUDA (kernelbench.com)Coding agents write four CUDA-only GPU kernels (GLM-5.2 fused MoE, DeepSeek sparse attention, MegaQwen decode, grid MinGRU scan) where Triton and other DSLs fail, one unlimited session per problem on an RTX PRO 6000, each run audited for reward hacking. Mean per-problem score over the full deck (peak fraction of roofline; sparse attention against a reference latency, so it can exceed 100%); failed, flagged or unaudited problems count 0. Independent of Stanford's KernelBench.
kernelbench.comTop model: Sonnet 5.522 models3 mixesQuality 60
- ClutterBench1 sub-boardClutter as a share of drawn area in generated graphics, using the raw prompt. Red area plus half the orange area, averaged over 6 scored posts. Lower is better.
clutterbench.adamholter.chatgpt.siteTop model: Sonnet 5.511 models7 mixesQuality 48
- DAYJOB: HealthcareUses Surge AI's rubric-graded healthcare assignment success rate. Score: score. Collected by Jake Boggs' benchmark dashboard.
surgehq.aiTop model: Opus 5.535 models1 mixQuality 70
- Legal Research BenchEvaluating agents on legal research tasks across diverse areas of US law
- Vending-Bench 2Year-long simulated vending-machine business; agents handle suppliers, inventory, pricing, and refunds. Score is the final bank balance.
- AA Intelligence Index v4.3.2Artificial Analysis' composite intelligence score (v4.3.2), combining 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1.