- Agentic terminal coding Terminal-Bench 2.0 59.1
- Agentic coding SWE-bench Verified 79.6
- Agentic computer use OSWorld-Verified 72.5
- Mean of Agentic tool use τ2-bench Retail / Agentic tool use τ2-bench Telecom 94.8
- Scaled tool use MCP-Atlas 61.3
- Agentic search BrowseComp 74.7
- Multidisciplinary reasoning Humanity's Last Exam (HLE) (with tools) 49
- Agentic financial analysis Finance Agent v1.1 63.3
- Office tasks GDPval-AA Elo 1633
- Novel problem-solving ARC-AGI-2 58.3
- Graduate-level reasoning GPQA Diamond 89.9
- Visual reasoning MMMU-Pro (with tools) 75.6
- Multilingual Q&A MMMLU 89.3
Claude Sonnet 4.6
Claude Sonnet 4.6 ranks first on 10 of the 297 benchmarks with five or more models scored.
309
Benchmarks with a score
10
Ranked first, out of five or more models
44
In the top three
Anthropic Feb 2026
Claude Sonnet 4.6
Closed weights Mid-size
152.9
#56 of 256
Epoch Capabilities Index
Self-Reported Index 152.8 #76 of 404
- SWE-bench Verified 75.2% #14 /32
- MMMU-Pro 75.6% #38 /90
- Terminal-Bench 2.0 53.4% #15 /43
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BullshitBench | 94.5% | #2 of 107 | Official leaderboard | |
| BullshitBench V2 | 91% | #3 of 129 | Official leaderboard | |
| Judgemark v2.1 | 94.1 | #1 of 42 | Official leaderboard | |
| Visual Aesthetic Benchmark | 40.3% | #1 of 18 | Official leaderboard | |
| EQ-Bench 3 | 1,710 | #5 of 78 | Official leaderboard | |
| MMLU-Pro | 87.3% | #14 of 212 | Official leaderboard | |
| Global PIQA | 90.2% | #4 of 44 | Official leaderboard | |
| OpenAI MRCR v2 | 72.7% | #5 of 47 | Official leaderboard | |
| LMArena (Elo) | 1,472 | #40 of 335 | Official leaderboard | |
| Judgemark v4 | 82.2 | #5 of 40 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 79.9 | #15 of 119 | Official leaderboard | |
| MMMLU | 89.3% | #12 of 71 | Lab model card | |
| HELM Arabic Enterprise | 73.7 | #4 of 23 | Official leaderboard | |
| Phare | 74% | #11 of 62 | Official leaderboard | |
| Continual Learning Bench | 0.20 | #1 of 5 | Official leaderboard | |
| Creative Writing v3 | 77.7 | #25 of 120 | Official leaderboard | |
| HELM Arabic | 81.5 | #6 of 28 | Official leaderboard | |
| SEA-HELM | 74.1 | #18 of 71 | Official leaderboard | |
| Document Arena | 1,478 | #10 of 39 | Official leaderboard | |
| Lech Mazur Creative Writing | 1.61 | #14 of 50 | Official leaderboard | |
| Humanity's Last Exam | 49% | #44 of 154 | Lab model card | |
| FACTS Search | 76.4% | #9 of 31 | Official leaderboard | |
| DesignArena ASCII Art | 1,230 | #23 of 77 | Official leaderboard | |
| Extended NYT Connections | 80.9% | #40 of 103 | Official leaderboard | |
| Structured Output Benchmark | 85.4% | #8 of 19 | Official leaderboard | |
| SimpleQA Verified | 30.9% | #30 of 56 | Official leaderboard | |
| KINA | 30% | #21 of 37 | Official leaderboard | |
| Antidote | 1,012 | #22 of 36 | Official leaderboard | |
| Hemingway-bench | 1,001 | #26 of 42 | Official leaderboard | |
| EQ-Bench 4 | 1,197 | #20 of 32 | Official leaderboard | |
| Global MMLU | 88.7% | #9 of 14 | Lab model card | |
| Vectara Hallucination Leaderboard (2025 dataset) | 10.6% | #67 of 104 | Official leaderboard | |
| AA-LCR Extended | 58.1% | #32 of 49 | Official leaderboard | |
| MILU | 89.6% | #8 of 11 | Lab model card | |
| HLE Extended | 10.9% | #25 of 31 | Official leaderboard | |
| SkateBench v2 | 15.1% | #20 of 21 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIME 2025 | 95.6% | #23 of 143 | Lab model card | |
| AstaBench | 54.5% | #3 of 13 | Official leaderboard | |
| ProofBench | 45% | #16 of 59 | Official leaderboard | |
| FrontierMath | 32.4% | #15 of 55 | Official leaderboard | |
| DTBench | 89.9% | #53 of 180 | Official leaderboard | |
| GPQA Diamond | 87.4% | #70 of 229 | Epoch AI run | |
| ARC-AGI-2 | 60.4% | #36 of 98 | Official leaderboard | |
| FrontierMath Tier 4 | 8.30% | #22 of 55 | Epoch AI run | |
| OTIS Mock AIME 2024-2025 | 85.8% | #86 of 212 | Epoch AI run | |
| ForecastBench | 0.13 | #34 of 82 | Official leaderboard | |
| ARC-AGI | 86.5% | #44 of 95 | Official leaderboard | |
| Verifying the Verifiers (Project APE CRED benchmark) | 83.5% | #11 of 23 | Official leaderboard | |
| FLTEval | 23.7 | #2 of 4 | Official leaderboard | |
| SpatialBench (LatchBio) | 44.2% | #9 of 14 | Official leaderboard | |
| BixBench3 | 15.3 | #18 of 20 | Official leaderboard | |
| scBench (LatchBio SingleCellBench) | 41.7% | #24 of 26 | Official leaderboard | |
| LatchBio Capabilities v1.0 | 31.5% | #25 of 27 | Official leaderboard | |
| BioMysteryBench | 50.4% | #10 of 10 | Lab model card | |
| Organic Chemistry (Anthropic) | 48.4% | #5 of 5 | Lab model card | |
| BioPipelineBench (Anthropic) | 52.1% | #3 of 3 | Lab model card | |
| Phylogenetics (Anthropic) | 49.1% | #3 of 3 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DesignArena Agentic Image to HTML | 1,247 | #1 of 11 | Official leaderboard | |
| SWE-bench Verified (hard subset) | 48.2% | #1 of 9 | Lab model card | |
| ts-bench | 80% | #1 of 7 | Official leaderboard | |
| Vibe Code Bench v1 | 29.1% | #3 of 17 | Official leaderboard | |
| DesignArena Website | 1,278 | #30 of 151 | Official leaderboard | |
| DesignArena Data Visualization | 1,276 | #26 of 118 | Official leaderboard | |
| Vision2Web S2-2026 | 41% | #2 of 9 | Official leaderboard | |
| WeirdML | 66.1% | #25 of 112 | Official leaderboard | |
| DesignArena Code | 1,266 | #39 of 143 | Official leaderboard | |
| ALE-Bench | 1,768 | #38 of 126 | Official leaderboard | |
| DesignArena UI Component | 1,263 | #44 of 139 | Official leaderboard | |
| DesignArena Game Dev | 1,249 | #49 of 143 | Official leaderboard | |
| WebDev Arena | 1,521 | #37 of 105 | arena.ai, via Epoch AI | |
| DesignArena React Native | 1,206 | #20 of 56 | Official leaderboard | |
| GBAEval | 48.8% | #8 of 21 | Official leaderboard | |
| DesignArena 3D Design | 1,217 | #52 of 136 | Official leaderboard | |
| DesignArena Full Stack | 1,191 | #19 of 49 | Official leaderboard | |
| SWE-bench Multilingual | 75.9% | #29 of 73 | Lab model card | |
| Code Migration | 39.9% | #31 of 78 | Official leaderboard | |
| Code Arena | 1,522 | #52 of 128 | Official leaderboard | |
| DesignArena SVG | 1,187 | #42 of 103 | Official leaderboard | |
| DesignArena Android (Kotlin) | 1,182 | #19 of 46 | Official leaderboard | |
| Text-based reinforcement learning task (Anthropic) | 0.93 | #3 of 7 | Lab model card | |
| SWE-bench Verified | 75.2% | #14 of 32 | Epoch AI run | |
| DesignArena Agentic Overall | 1,179 | #34 of 74 | Official leaderboard | |
| Vibe Code Bench | 55.8% | #62 of 120 | Official leaderboard | |
| Time series forecasting task (Anthropic) | 5.70 | #5 of 9 | Lab model card | |
| FrontierCode | 15.1% | #7 of 12 | Official leaderboard | |
| AndroidBench | 67% | #19 of 32 | Official leaderboard | |
| Image-to-WebDev Arena | 1,530 | #33 of 55 | Official leaderboard | |
| ProgramBench | 0% | #15 of 25 | Official leaderboard | |
| DesignArena Frontend Web App | 1,172 | #33 of 53 | Official leaderboard | |
| PostTrainBench | 16.4% | #15 of 23 | Official leaderboard | |
| SWE Atlas Codebase QnA Extended | 6.70% | #25 of 38 | Official leaderboard | |
| DesignArena Agentic Game Dev | 1,119 | #18 of 27 | Official leaderboard | |
| SWE Atlas | 31.2% | #15 of 22 | Official leaderboard | |
| SWE Atlas Test Writing | 31.8% | #15 of 22 | Official leaderboard | |
| APEX-SWE | 34.5% | #45 of 65 | Official leaderboard | |
| OpenHands Index | 44.5% | #23 of 33 | Official leaderboard | |
| DesignArena Image to Frontend | 1,147 | #22 of 31 | Official leaderboard | |
| Vision2Web S3-2026 | 58.9% | #5 of 7 | Official leaderboard | |
| ProgramBench tests passed | 47.5% | #18 of 25 | Official leaderboard | |
| Data Agent Benchmark | 68.2% | #13 of 18 | Official leaderboard | |
| SlopCodeBench | 16.8% | #16 of 22 | Official leaderboard | |
| ProgramBench almost solved | 1.00% | #19 of 25 | Official leaderboard | |
| SWE Atlas Refactoring | 32.2% | #13 of 17 | Official leaderboard | |
| SWE-bench Verified Extended | 27.3% | #36 of 44 | Official leaderboard | |
| DeepSWE | 29.9% | #27 of 28 | Official leaderboard | |
| FrontierCode 1.1 | 24.3% | #8 of 8 | Official leaderboard | |
| AgenticDataBench | 46.7% | #3 of 3 | Official leaderboard | |
| OpenRCA | 27.9% | #2 of 2 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Claw-Eval | 81.4% | #1 of 24 | Official leaderboard | |
| QwenClawBench | 60.5% | #1 of 12 | Official leaderboard | |
| tau2-bench | 94.8% | #7 of 79 | Lab model card | |
| LisanBench | 12060.0 | #12 of 133 | Official leaderboard | |
| DeepResearch Bench | 54.9% | #2 of 22 | evals.futuresearch.ai, via Epoch AI | |
| Deep Research Bench (FutureSearch) | 0.55 | #2 of 21 | Official leaderboard | |
| OSWorld-Verified | 81.5% | #4 of 33 | Official leaderboard | |
| Claw-Eval Pass^3 | 67.8% | #3 of 24 | Official leaderboard | |
| OSWorld | 72.1% | #1 of 8 | OS World Website, via Epoch AI | |
| EnterpriseOps-Gym | 42.2% | #2 of 12 | Official leaderboard | |
| PinchBench | 76.4% | #9 of 48 | Official leaderboard | |
| UniClawBench | 45.5% | #2 of 10 | Official leaderboard | |
| Pencil Puzzle Bench | 26.7% | #11 of 53 | Official leaderboard | |
| WebArena | 65.6% | #2 of 8 | Lab model card | |
| Toolathlon | 44.8% | #13 of 50 | Official leaderboard | |
| OpenComputer | 64.4% | #2 of 7 | Official leaderboard | |
| SnakeBench | 26 | #61 of 188 | Official leaderboard | |
| Claw-Eval-Live | 61.9% | #4 of 12 | Official leaderboard | |
| ClawBench | 56.9% | #2 of 6 | Official leaderboard | |
| CocoaBench | 34% | #2 of 6 | Official leaderboard | |
| iOSWorld | 46.6% | #2 of 6 | Official leaderboard | |
| Terminal-Bench 2.0 | 53.4% | #15 of 43 | Official leaderboard | |
| PACT | 1,546 | #10 of 26 | Official leaderboard | |
| AutoResearchBench | 6.96% | #5 of 13 | Official leaderboard | |
| ClawProBench | 60.5% | #19 of 49 | Official leaderboard | |
| LMArena Search Arena | 1,201 | #16 of 37 | Official leaderboard | |
| DeepSearchQA | 89.2% | #10 of 23 | Lab model card | |
| LLM Chess | 190 | #55 of 123 | Official leaderboard | |
| Clocktower Radio | 1,615 | #11 of 23 | Official leaderboard | |
| MM-ClawBench | 64.2% | #3 of 6 | Official leaderboard | |
| GAIA | 54.5% | #33 of 64 | Official leaderboard | |
| Long-Horizon-Terminal-Bench (LHTB) | 37.3% | #13 of 25 | Official leaderboard | |
| BrowseComp | 74.7% | #43 of 81 | Lab model card | |
| MyPCBench | 50.5% | #5 of 9 | Official leaderboard | |
| Game Arena Clobber | 262 | #22 of 39 | Official leaderboard | |
| OSWorld 2.0 | 9.30% | #4 of 7 | Official leaderboard | |
| WebArena-Verified | 66.4% | #4 of 7 | Lab model card | |
| Game Arena Word Association | 216 | #23 of 40 | Official leaderboard | |
| Game Arena Werewolf | -2 | #5 of 8 | Official leaderboard | |
| Chess Puzzles | 23% | #59 of 93 | Epoch AI run | |
| GameWorld | 39.3% | #14 of 22 | Official leaderboard | |
| Game Arena Dots and Boxes | 801 | #26 of 40 | Official leaderboard | |
| Game Arena Coin Game | 293 | #26 of 39 | Official leaderboard | |
| CAR-bench | 50% | #8 of 12 | Official leaderboard | |
| Era by Eon | 63.6% | #6 of 9 | Official leaderboard | |
| SkillsBench | 47.2% | #15 of 22 | Official leaderboard | |
| CEO-Bench | $38,166 | #13 of 19 | Official leaderboard | |
| PinchBench 2.0 | 62.7% | #40 of 57 | Official leaderboard | |
| MCP Atlas | 69.5% | #26 of 37 | Official leaderboard | |
| Game Arena Checkers | 689 | #29 of 41 | Official leaderboard | |
| BrowseComp Extended | 14.2% | #30 of 42 | Official leaderboard | |
| Gaia2-CLI | 51.9% | #5 of 7 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 186 | #28 of 39 | Official leaderboard | |
| Kaggle Game Arena | 203 | #31 of 42 | Official leaderboard | |
| TUA-Bench | 42.8% | #9 of 12 | Official leaderboard | |
| CUA-World-Long | 6.00% | #3 of 4 | Official leaderboard | |
| Mystery Game Puzzles | 16% | #64 of 84 | Epoch AI run | |
| AutomationBench | 5.33% | #32 of 42 | Official leaderboard | |
| Game Arena Reversi | 215 | #30 of 39 | Official leaderboard | |
| Game Arena Nine Men's Morris | 428 | #31 of 39 | Official leaderboard | |
| Game Arena Word Art | 218 | #31 of 39 | Official leaderboard | |
| Game Arena Dark Hex | 115 | #34 of 41 | Official leaderboard | |
| Game Arena Lines of Action | 209 | #33 of 39 | Official leaderboard | |
| EnvSimBench | 37.8% | #6 of 7 | Official leaderboard | |
| Terminal-Bench 2.1 Extended | 17.5% | #38 of 44 | Official leaderboard | |
| RuneBench | 3.18 | #66 of 76 | Official leaderboard | |
| Game Arena Five in a Row | 241 | #18 of 20 | Official leaderboard | |
| Game Arena Poker | -19.86 | #28 of 31 | Official leaderboard | |
| Game Arena Four in a Row | 133 | #39 of 43 | Official leaderboard | |
| Game Arena Go | 296 | #36 of 39 | Official leaderboard | |
| Game Arena Chess | 98 | #39 of 42 | Official leaderboard | |
| Game Arena Bargaining | 75 | #38 of 39 | Official leaderboard | |
| ApprenticeBench | 2.00% | #18 of 18 | Official leaderboard | |
| Toloka Arena | 21.6% | #5 of 5 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| TaxEval v2 | 77.1% | #4 of 129 | Official leaderboard | |
| Vals Finance Agent | 63.3% | #2 of 49 | Official leaderboard | |
| Vals Index v1 | 65.5% | #5 of 34 | Official leaderboard | |
| CHI-Bench | 26.2% | #4 of 23 | Official leaderboard | |
| ProfBench | 55.6% | #15 of 72 | Official leaderboard | |
| EnterpriseClawBench | 64.4% | #2 of 9 | Official leaderboard | |
| MortgageTax | 67.7% | #21 of 87 | Official leaderboard | |
| CorpFin v2 | 65.3% | #31 of 119 | Official leaderboard | |
| CaseLaw v2 | 64% | #14 of 53 | Official leaderboard | |
| GDPval-AA | 1,633 | #5 of 16 | Lab model card | |
| Vending-Bench 2 | $7,204 | #21 of 67 | Official leaderboard | |
| RCM-Bench-1 | 89.8% | #8 of 20 | Official leaderboard | |
| TaxCalcBench | 37.3% | #7 of 17 | Official leaderboard | |
| Legal Research Bench | 38.5% | #40 of 93 | Official leaderboard | |
| Excel Modeling Benchmark | 60.1% | #34 of 76 | Official leaderboard | |
| PhysicianBench | 23% | #6 of 13 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 5.00% | #37 of 80 | Official leaderboard | |
| Vending-Bench Arena | $5,639 | #15 of 32 | Official leaderboard | |
| Vals Finance Agent v2 | 51% | #41 of 80 | Official leaderboard | |
| Tax Agent Bench | 25.8% | #44 of 84 | Official leaderboard | |
| LegalBench | 82.1% | #73 of 138 | Official leaderboard | |
| Public Benefits Bench | 62.5% | #27 of 50 | Official leaderboard | |
| Vals Index v2 | 50.6% | #36 of 65 | Official leaderboard | |
| FoodTruck Bench | $17,426 | #33 of 59 | Official leaderboard | |
| 3DCodeArena | 983 | #8 of 14 | Official leaderboard | |
| OccuBench | 64.4% | #8 of 14 | Official leaderboard | |
| GDPval Extended | 0.80% | #26 of 43 | Official leaderboard | |
| FORTE | 68.3% | #13 of 21 | Official leaderboard | |
| JobBench | 36.6% | #11 of 16 | Official leaderboard | |
| PolicyBench | 79.9% | #35 of 50 | Official leaderboard | |
| CADGenBench | 28.3% | #8 of 11 | Official leaderboard | |
| TRLawBench | 8.38 | #25 of 34 | Official leaderboard | |
| APEX-Agents | 43% | #46 of 61 | Collected by Epoch AI, source not named | |
| QF-Bench | 56.3% | #7 of 9 | Official leaderboard | |
| APEX-Agents 1.1 | 57.6% | #51 of 65 | Official leaderboard | |
| APEX-1 | 55.1% | #62 of 71 | Official leaderboard | |
| EEBench | 20.4% | #27 of 30 | Official leaderboard | |
| YC-Bench | $104,432 | #55 of 59 | Official leaderboard | |
| MedXpertQA MM Extended | 41.2% | #39 of 41 | Official leaderboard | |
| prinzbench | 8.00 | #26 of 26 | Official leaderboard | |
| Real-World Finance (Anthropic) | 59.5% | #3 of 3 | Lab model card | |
| MedCalc-Bench Verified | 86.2% | #1 of 1 | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vision Arena | 1,275 | #27 of 136 | Official leaderboard | |
| MeasureBench Synthetic | 18.5% | #4 of 19 | Official leaderboard | |
| MeasureBench Real-world | 20.1% | #5 of 19 | Official leaderboard | |
| vCX Benchmark | 89.5% | #15 of 42 | Official leaderboard | |
| Claw-Eval-MM | 50.9% | #4 of 11 | Official leaderboard | |
| AgentWorldBench | 56.0 | #5 of 13 | Official leaderboard | |
| MMMU-Pro | 75.6% | #38 of 90 | Lab model card | |
| VisFactor | 24.7% | #11 of 26 | Official leaderboard | |
| ZeroBench | 12% | #35 of 76 | Official leaderboard | |
| GeoBench | 73% | #24 of 48 | Official leaderboard | |
| BVB | 65.3 | #26 of 45 | Official leaderboard | |
| MDPBench | 73.1 | #7 of 12 | Official leaderboard | |
| MineBench | 1,345 | #47 of 75 | Official leaderboard | |
| CC-OCR-Bench-V2 | 59.8% | #10 of 15 | Official leaderboard | |
| CharXiv | 72.4% | #52 of 75 | Lab model card | |
| 1H-VideoQA | 67.3% | #14 of 20 | Official leaderboard | |
| Blueprint-Bench 2 | 6.70% | #32 of 38 | Official leaderboard | |
| AgenticVBench | 16.6% | #8 of 9 | Official leaderboard | |
| CharXiv Extended | 64.4% | #37 of 38 | Official leaderboard | |
| MMMU-Pro Extended | 19.9% | #37 of 38 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveSecBench | 86.0 | #2 of 34 | Official leaderboard | |
| Malicious computer use (Anthropic) | 99.4% | #1 of 14 | Lab model card | |
| KORA | 67.1% | #4 of 42 | Official leaderboard | |
| BioTIER Refuse | 95.2% | #8 of 69 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, coding (Anthropic) | 0% | #2 of 14 | Lab model card | |
| EvalDetectBench | 0.88 | #2 of 10 | Official leaderboard | |
| nocot-bench (NCRI) | 87.8 | #38 of 185 | Official leaderboard | |
| Single-turn harmful request evaluations (Anthropic) | 99.4% | #4 of 19 | Lab model card | |
| Anthropic political even-handedness eval | 98.4% | #3 of 14 | Lab model card | |
| World-Class Bio | 55.4% | #17 of 74 | Official leaderboard | |
| Higher-difficulty harmful request evaluations (Anthropic) | 99.4% | #1 of 4 | Lab model card | |
| Cybench | 90% | #5 of 18 | Lab model card | |
| Creative biology | 59.3% | #2 of 7 | Lab model card | |
| HarnessAudit-Bench | 0.32 | #2 of 7 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 0.24% | #4 of 13 | Lab model card | |
| HelpBench | 83% | #3 of 9 | Official leaderboard | |
| Pwn CTF challenges (Anthropic) | 71.4% | #2 of 6 | Lab model card | |
| Web CTF challenges (Anthropic) | 100% | #2 of 6 | Lab model card | |
| Impossible tasks (Anthropic) | 40% | #4 of 11 | Lab model card | |
| Reward-hack-prone coding tasks (Anthropic) | 0% | #3 of 8 | Lab model card | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 60% | #28 of 71 | Official leaderboard | |
| BBQ (Bias Benchmark for QA) | 88.1% | #10 of 25 | Lab model card | |
| SMDD-Bench | 38% | #3 of 7 | Official leaderboard | |
| ExploitBench | 26% | #4 of 9 | Official leaderboard | |
| LAB-Bench FigQA | 58.8% | #5 of 11 | Lab model card | |
| CyberGym | 65.2% | #16 of 34 | Official leaderboard | |
| BiomniBench | 62.4 | #4 of 8 | Official leaderboard | |
| Crypto CTF challenges (Anthropic) | 88.9% | #3 of 6 | Lab model card | |
| Network CTF challenges (Anthropic) | 100% | #3 of 6 | Lab model card | |
| Rev CTF challenges (Anthropic) | 100% | #3 of 6 | Lab model card | |
| Participation in junk science (Anthropic) | 0% | #1 of 2 | Lab model card | |
| Gray Swan Agent Red Teaming (ART) benchmark | 2.20% | #8 of 14 | Lab model card | |
| Agentic coding malicious requests (Anthropic) | 100% | #3 of 5 | Lab model card | |
| VariantBench | 31.1% | #19 of 27 | Official leaderboard | |
| MetagenomicsBench | 28.3% | #19 of 26 | Official leaderboard | |
| SHADE-Arena | 1.60% | #12 of 16 | Lab model card | |
| DrugDiscoveryBench | 31.3% | #16 of 20 | Official leaderboard | |
| Over-eager behavior on GUI computer use tasks (Anthropic) | 14.7% | #4 of 5 | Lab model card | |
| SpatialBench-Long | 22.6% | #22 of 27 | Official leaderboard | |
| TxBench-Oligo-Discovery | 30.1% | #25 of 30 | Official leaderboard | |
| EpiBench | 18.9% | #23 of 27 | Official leaderboard | |
| scBench-Long | 10.2% | #23 of 27 | Official leaderboard | |
| TxBench-Antibody-Discovery | 26.3% | #26 of 30 | Official leaderboard | |
| Malicious use of Claude Code (Anthropic) | 79.3% | #15 of 17 | Lab model card | |
| BioSecBench-Function | 27.3% | #24 of 27 | Official leaderboard | |
| BioSecBench-Surveillance | 34.5% | #24 of 27 | Official leaderboard | |
| Single-turn benign request evaluations (Anthropic) | 0.41% | #17 of 19 | Lab model card | |
| OncologyBench | 5.33% | #27 of 30 | Official leaderboard | |
| Virology Capabilities Test (VCT) | 39.7% | #28 of 31 | Lab model card | |
| BioTIER Permit | 81.6% | #64 of 69 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, computer use (Anthropic) | 12% | #13 of 14 | Lab model card | |
| TxBench-Preclinical-Pharmacology | 34.3% | #28 of 30 | Official leaderboard | |
| Agentic BAIM-LLM Evaluation (ABLE) | 0% | #17 of 18 | Official leaderboard | |
| DNA Synthesis Screening Evasion | 30% | #5 of 5 | Lab model card | |
| Refusal to assist with AI safety R&D (Anthropic) | 38% | #5 of 5 | Lab model card | |
| Higher-difficulty benign request evaluations (Anthropic) | 0.18% | #4 of 4 | Lab model card | |
| SSH stress-testing evaluation (Anthropic) | 92% | #2 of 2 | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ACCoRD | 73.4 | #34 of 152 | Official leaderboard | |
| CAIS Risk Index | 40.4 | #16 of 67 | Official leaderboard | |
| LMCA | 46.5% | #37 of 147 | Official leaderboard | |
| Conceptual Reasoning Index | 59.2 | #40 of 151 | Official leaderboard | |
| SAGE (Vals) | 46.6% | #34 of 81 | Official leaderboard | |
| Vals Multimodal Index | 60.6% | #17 of 33 | Official leaderboard | |
| CAIS Text Capabilities Index | 26.3 | #42 of 68 | Official leaderboard | |
| CAIS Vision Capabilities Index | 46.3 | #36 of 53 | Official leaderboard |
Its lab's numbers put it at 152.8, in line with its ECI
The Self-Reported Index fits one score to the 21 capability results Anthropic reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 150.7 to 154.8. How we compute it.
- AIME 2025 95.6
- tau2-bench 94.8
- GPQA Diamond 89.9
- MILU 89.6
- MMMLU 89.3
- DeepSearchQA 89.2
- Global MMLU 88.7
- ARC-AGI 86.5
- SWE-bench Verified 79.6
- SWE-bench Multilingual 75.9
- MMMU-Pro 75.6
- BrowseComp 74.7
- OSWorld-Verified 72.5
- CharXiv 72.4
- Vals Finance Agent 63.3
- MCP Atlas 61.3
- Terminal-Bench 2.0 59.1
- ARC-AGI-2 58.3
- BioMysteryBench 50.4
- Humanity's Last Exam 49
- SWE-bench Verified (hard subset) 48.2
Its lab's numbers next to 11 leaderboards
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| OSWorld-Verified | 72.5% Agentic computer use OSWorld-Verified | 81.5% Leaderboard, 100 steps, Pointer Agent | −9.0 pts |
| MCP Atlas | 61.3% Scaled tool use MCP-Atlas | 69.5% Leaderboard | −8.2 pts |
| Terminal-Bench 2.0 | 59.1% Agentic terminal coding Terminal-Bench 2.0 | 53.4% Leaderboard, Simplai Agent | +5.7 pts |
| SWE-bench Verified | 79.6% Agentic coding SWE-bench Verified | 75.2% Epoch AI run, no thinking | +4.4 pts |
| GPQA Diamond | 89.9% Graduate-level reasoning GPQA Diamond | 87.4% Epoch AI run, 32k thinking | +2.5 pts |
| ARC-AGI-2 | 58.3% Novel problem-solving ARC-AGI-2 | 60.4% Leaderboard, high | −2.1 pts |
| Vals Finance Agent | 63.3% Agentic financial analysis Finance Agent v1.1 | 63.3% Leaderboard | Same |
| ARC-AGI | 86.5% ARC-AGI-1, 120k thinking tokens, high effort, ARC Prize private set (2.7, Figure 2.7.A) | 86.5% Leaderboard, high | Same |
| CyberGym | 65.2% CyberGym targeted vulnerability reproduction, pass@1 over 1,507 tasks, no thinking with think tool (2.15) | 65.2% Leaderboard, Anthropic Agent | Same |
| MMMU-Pro | 75.6% Visual reasoning MMMU-Pro (with tools) | 75.6% Leaderboard, with tools | Same |
| Vending-Bench 2 | $7,204 Vending-Bench 2 final balance, max effort; $6,625.10 at high effort (2.13) | $7,204 Leaderboard | Same |
Anthropic's models over time
Anthropic's models in color, against every other model on the Epoch Capabilities Index, by release date.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 67 of them aren't in the launch post.
System Card: Claude Sonnet 4.6 Feb 17, 2026
Capabilities
- SWE-bench Verified 79.6
- Terminal-Bench 2.0 59.1
- tau2-bench
- MCP Atlas 61.3
- OSWorld-Verified 72.5
- ARC-AGI-2 58.3
- GPQA Diamond 89.9
- MMMLU 89.3
- GDPval-AA 1633
- MMMU-Pro 74.5
- Humanity's Last Exam 33.2
- SWE-bench Multilingual 75.9
- OpenRCA 27.9
- ARC-AGI 86.5
- AIME 2025 95.6
- Vals Finance Agent 63.3
- Real-World Finance (Anthropic) 59.5
- Vending-Bench 2 7204.1
- CyberGym 65.2
- OpenAI MRCR v2
- GraphWalks
- LAB-Bench FigQA 58.8
- CharXiv 72.4
- WebArena 65.6
- WebArena-Verified 66.4
- Global MMLU 88.7
- MILU 89.6
- BrowseComp 74
- DeepSearchQA 89.2
- BioPipelineBench (Anthropic) 52.1
- BioMysteryBench 50.4
- Structural Biology (Anthropic)
- Organic Chemistry (Anthropic) 48.4
- Phylogenetics (Anthropic) 49.1
Health
Safety
- Single-turn harmful request evaluations (Anthropic) 99.4
- Single-turn benign request evaluations (Anthropic) 0.41
- Higher-difficulty harmful request evaluations (Anthropic) 99.4
- Higher-difficulty benign request evaluations (Anthropic) 0.18
- Multi-turn testing (Anthropic)
- Child safety evaluations (Anthropic)
- Suicide and self-harm evaluations (Anthropic)
- SSH stress-testing evaluation (Anthropic) 92
- Agentic coding malicious requests (Anthropic) 100
- Malicious use of Claude Code (Anthropic) 79.3
- Malicious computer use (Anthropic) 99.4
Robustness
Alignment
- Reward-hack-prone coding tasks (Anthropic) 0
- Impossible tasks (Anthropic) 40
- Agentic code behavior scores (Anthropic)
- Over-eager behavior on GUI computer use tasks (Anthropic) 14.7
- Anthropic automated behavioral audit
- Petri
- Refusal to assist with AI safety R&D (Anthropic) 38
- Self-preference bias (Anthropic)
- Participation in junk science (Anthropic) 0
- SHADE-Arena 1.6