- Agentic coding SWE-bench Pro 64.3
- Agentic coding SWE-bench Verified 87.6
- SWE-bench Multilingual 80.5
- SWE-bench Multimodal 34.5
- Agentic terminal coding Terminal-Bench 2.0 69.4
- Multidisciplinary reasoning Humanity's Last Exam (with tools) 54.7
- Agentic search BrowseComp 79.3
- Scaled tool use MCP-Atlas 77.3
- Agentic computer use OSWorld-Verified 78
- Agentic financial analysis Finance Agent v1.1 64.4
- Graduate-level reasoning GPQA Diamond 94.2
- Visual reasoning CharXiv Reasoning (with tools) 91
- Multilingual Q&A MMMLU 91.5
- ScreenSpot-Pro (with tools) 87.6
- OfficeQA 86.3
- OfficeQA Pro 80.6
- ARC-AGI-1 92
- ARC-AGI-2 75.8
- CursorBench, 70% at high effort 70
- BigLaw Bench, 90.9% at high effort 90.9
- XBOW visual-acuity benchmark, 98.5% 98.5
Claude Opus 4.7
Claude Opus 4.7 ranks first on 20 of the 290 benchmarks with five or more models scored.
311
Benchmarks with a score
20
Ranked first, out of five or more models
61
In the top three
Anthropic Apr 2026
Claude Opus 4.7
Closed weights Flagship
157.4
#30 of 256
Epoch Capabilities Index
Self-Reported Index 159.6 #34 of 404
- Humanity's Last Exam 36.2% #10 /47
- SWE-bench Verified 83.5% #1 /32
- Terminal-Bench 2.0 80.2% #3 /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 | |
|---|---|---|---|---|
| LMArena (Elo) | 1,501 | #8 of 335 | Official leaderboard | |
| HELM Arabic | 84.1 | #1 of 28 | Official leaderboard | |
| EQ-Bench 3 | 1,914 | #3 of 78 | Official leaderboard | |
| HELM Arabic Enterprise | 77.9 | #1 of 23 | Official leaderboard | |
| BullshitBench V2 | 83% | #8 of 129 | Official leaderboard | |
| BullshitBench | 80% | #7 of 107 | Official leaderboard | |
| MMMLU | 91.5% | #5 of 71 | Lab model card | |
| Creative Writing v3 | 80.7 | #9 of 120 | Official leaderboard | |
| Judgemark v4 | 84.0 | #3 of 40 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 81.8 | #9 of 119 | Official leaderboard | |
| INCLUDE | 87% | #3 of 31 | Lab model card | |
| Document Arena | 1,494 | #4 of 39 | Official leaderboard | |
| Corp Speak Arena | 1,644 | #3 of 22 | Official leaderboard | |
| CulturalRiddles | 66% | #3 of 20 | Official leaderboard | |
| FACTS Search | 78.8% | #5 of 31 | Official leaderboard | |
| HUMAINE | 32 | #10 of 50 | Official leaderboard | |
| Structured Output Benchmark | 86.4% | #4 of 19 | Official leaderboard | |
| Humanity's Last Exam | 36.2% | #10 of 47 | Official leaderboard | |
| EQ-Bench 4 | 1,293 | #7 of 32 | Official leaderboard | |
| Lech Mazur Creative Writing | 1.72 | #13 of 50 | Official leaderboard | |
| SimpleQA Verified | 51.7% | #29 of 94 | Epoch AI run | |
| Antidote | 1,042 | #14 of 36 | Official leaderboard | |
| Hemingway-bench | 1,042 | #17 of 42 | Official leaderboard | |
| Global MMLU | 89.9% | #7 of 14 | Lab model card | |
| xDailyBench | 69% | #2 of 4 | Official leaderboard | |
| HLE Extended | 31.2% | #17 of 31 | Official leaderboard | |
| FACTS Parametric | 38.8% | #25 of 45 | Official leaderboard | |
| Continual Learning Bench | 0.18 | #3 of 5 | Official leaderboard | |
| AA-LCR Extended | 59.8% | #31 of 49 | Official leaderboard | |
| MILU | 89.9% | #7 of 11 | Lab model card | |
| Joke Arena | 1,489 | #16 of 25 | Official leaderboard | |
| False-premise questions (Anthropic) | 77.2% | #4 of 6 | Lab model card | |
| Extended NYT Connections | 39% | #77 of 103 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 12% | #80 of 104 | Official leaderboard | |
| OpenAI MRCR v2 | 2.74% | #47 of 47 | Official leaderboard | |
| 100Q-Hard (Anthropic) | 17.1% | #7 of 7 | Lab model card | |
| Doubao Multi-Turn Bench | 49.8% | #4 of 4 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ChemBench | 70.9% | #1 of 24 | Official leaderboard | |
| FrontierMath | 43.8% | #4 of 55 | Official leaderboard | |
| AstaBench | 58% | #1 of 13 | Official leaderboard | |
| FrontierMath Tier 4 | 22.9% | #8 of 55 | Epoch AI run | |
| MathArena (Apex, IMO, USAMO) | 40.6% | #7 of 47 | Official leaderboard | |
| DTBench | 94.7% | #28 of 180 | Official leaderboard | |
| ProofBench | 54% | #10 of 59 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 97.8% | #36 of 212 | Epoch AI run | |
| GPQA Diamond | 90.2% | #46 of 229 | Epoch AI run | |
| xbench-ScienceQA | 73% | #6 of 29 | Official leaderboard | |
| NatureBench | 15.6% | #4 of 18 | Official leaderboard | |
| ARC-AGI-2 | 75.8% | #23 of 98 | Official leaderboard | |
| ARC-AGI | 93.5% | #27 of 95 | Official leaderboard | |
| ResearchClawBench | 20.7 | #7 of 23 | Official leaderboard | |
| HMMT 2026 | 93.9% | #10 of 31 | Official leaderboard | |
| Phylogenetics (Anthropic) | 79.6% | #1 of 3 | Lab model card | |
| FrontierMath Tiers 1-3 v2 | 70.2% | #29 of 86 | Official leaderboard | |
| SimpleBench | 61.7% | #36 of 106 | Official leaderboard | |
| SpatialBench (LatchBio) | 52.4% | #6 of 14 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 31.7% | #31 of 68 | Official leaderboard | |
| AIME 2026 | 95.8% | #15 of 31 | Official leaderboard | |
| MathArena Apex Shortlist | 63.8% | #21 of 38 | Official leaderboard | |
| Organic Chemistry (Anthropic) | 77.2% | #3 of 5 | Lab model card | |
| ForecastBench | 0.14 | #50 of 82 | Official leaderboard | |
| LatchBio Capabilities v1.0 | 38.6% | #18 of 27 | Official leaderboard | |
| scBench (LatchBio SingleCellBench) | 52.5% | #19 of 26 | Official leaderboard | |
| Leipzig Benchmark | 13% | #15 of 19 | Official leaderboard | |
| USAMO 2026 | 69.3% | #6 of 6 | Lab model card | |
| BioPipelineBench Verified (Anthropic) | 83.6% | #1 of 1 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-bench Verified | 83.5% | #1 of 32 | Epoch AI run | |
| FeatureBench | 78.2% | #1 of 16 | Official leaderboard | |
| FeatureBench Lite | 78.2% | #1 of 16 | Official leaderboard | |
| EvoCode-Bench | 54% | #1 of 13 | Official leaderboard | |
| OpenHands Index | 69.7% | #3 of 33 | Official leaderboard | |
| RoadmapBench | 39.1% | #1 of 11 | Official leaderboard | |
| WeirdML | 76.4% | #16 of 112 | Official leaderboard | |
| PostTrainBench | 28.6% | #4 of 23 | Official leaderboard | |
| IOI (Vals) | 47.1% | #10 of 55 | Official leaderboard | |
| Claw-SWE-Bench | 77.1% | #3 of 16 | Official leaderboard | |
| SWE-bench Multilingual | 80.5% | #14 of 73 | Lab model card | |
| GSO-Bench | 44.1% | #6 of 30 | Official leaderboard | |
| ALE-Bench | 1,897 | #27 of 126 | Official leaderboard | |
| SWE-Bench Pro | 64.3% | #18 of 81 | Lab model card | |
| Time series forecasting task (Anthropic) | 4.78 | #2 of 9 | Lab model card | |
| SWE Atlas Refactoring | 48.6% | #4 of 17 | Official leaderboard | |
| FrontierCode | 23% | #3 of 12 | Official leaderboard | |
| WebDev Arena | 1,567 | #28 of 105 | arena.ai, via Epoch AI | |
| APEX-SWE | 47.6% | #20 of 65 | Official leaderboard | |
| Code Migration | 43.9% | #26 of 78 | Official leaderboard | |
| SWE-Marathon | 16% | #4 of 12 | Official leaderboard | |
| Code Arena | 1,557 | #43 of 128 | Official leaderboard | |
| Quadruped reinforcement learning task (Anthropic) | 26.5 | #3 of 8 | Lab model card | |
| Data Agent Benchmark | 81.8% | #7 of 18 | Official leaderboard | |
| Vibe Code Bench | 71% | #48 of 120 | Official leaderboard | |
| Image-to-WebDev Arena | 1,579 | #22 of 55 | Official leaderboard | |
| TasteVal | 50.1× | #8 of 20 | Official leaderboard | |
| CursorBench 3.1 | 70% | #2 of 5 | Lab model card | |
| Vision2Web S3-2026 | 65.9% | #3 of 7 | Official leaderboard | |
| SWE-bench Verified Extended | 58% | #20 of 44 | Official leaderboard | |
| GameCraft-Bench | 41.5% | #7 of 15 | Official leaderboard | |
| FrontierSWE v1 | 56.3% | #8 of 17 | Official leaderboard | |
| SWE Atlas | 40.3% | #11 of 22 | Official leaderboard | |
| SWE-Together | 61.5% | #11 of 22 | Official leaderboard | |
| FrontierCS | 43.0 | #7 of 14 | Official leaderboard | |
| EvoPolicyGym | 0.75 | #2 of 4 | Official leaderboard | |
| ParallelKernelBench | 23% | #2 of 4 | Official leaderboard | |
| ProgramBench almost solved | 4.50% | #13 of 25 | Official leaderboard | |
| AndroidBench | 68.7% | #17 of 32 | Official leaderboard | |
| GBAEval | 43.8% | #12 of 21 | Official leaderboard | |
| SWE Atlas Codebase QnA Extended | 11.3% | #22 of 38 | Official leaderboard | |
| SlopCodeBench | 20.9% | #13 of 22 | Official leaderboard | |
| SWE Atlas Test Writing | 38.5% | #13 of 22 | Official leaderboard | |
| ProgramBench tests passed | 55.1% | #15 of 25 | Official leaderboard | |
| FrontierCode 1.1 | 38.5% | #5 of 8 | Official leaderboard | |
| PostTrainBench 1.1 | 28.6% | #7 of 11 | Official leaderboard | |
| FrogsGame post-training task | 1.00% | #11 of 17 | Official leaderboard | |
| ProgramBench | 0% | #19 of 25 | Official leaderboard | |
| PostTrainBench 1.2 | 24.3% | #16 of 21 | Official leaderboard | |
| MirrorCode | 31.1% | #30 of 39 | Epoch AI run | |
| Vals RSI Index | 0.19 | #21 of 27 | Official leaderboard | |
| Senior SWE-Bench | 18.1% | #18 of 23 | Official leaderboard | |
| Doc2Harness | 32.2% | #15 of 19 | Official leaderboard | |
| SWE-bench Multimodal | 34.5% | #10 of 12 | Lab model card | |
| Novel compiler task (Anthropic) | 71.1% | #4 of 4 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LisanBench | 14400.0 | #7 of 133 | Official leaderboard | |
| OSWorld-Verified | 83.6% | #2 of 33 | Official leaderboard | |
| Terminal-Bench 2.0 | 80.2% | #3 of 43 | Official leaderboard | |
| Toolathlon | 52.8% | #4 of 50 | Official leaderboard | |
| SkillsBench | 61.2% | #2 of 22 | Official leaderboard | |
| ScreenSpot-Pro | 87.6% | #3 of 32 | Lab model card | |
| Pencil Puzzle Bench | 50% | #5 of 53 | Official leaderboard | |
| WildClawBench | 62.2% | #3 of 31 | Official leaderboard | |
| DeepTerminalBench | 34% | #1 of 9 | Official leaderboard | |
| WeaveBench | 41.2% | #1 of 9 | Official leaderboard | |
| ClawBench V2 | 44.6% | #1 of 7 | Official leaderboard | |
| SnakeBench | 30 | #41 of 188 | Official leaderboard | |
| AppWorld Test-Challenge | 85.6% | #3 of 12 | Official leaderboard | |
| TUA-Bench | 58% | #3 of 12 | Official leaderboard | |
| TERMS-Bench | 0.66 | #3 of 11 | Official leaderboard | |
| MobileWorld | 45.8% | #5 of 18 | Official leaderboard | |
| τ³-bench Banking | 40.2% | #7 of 25 | Official leaderboard | |
| ADeptS-Bench Desktop | 75.4% | #2 of 7 | Official leaderboard | |
| ADeptS-Bench Mobile | 71.6% | #2 of 7 | Official leaderboard | |
| OSWorld 2.0 | 18.2% | #2 of 7 | Official leaderboard | |
| LLM Chess | 691 | #36 of 123 | Official leaderboard | |
| AppWorld Test-Normal | 88.7% | #4 of 12 | Official leaderboard | |
| LMArena Search Arena | 1,210 | #13 of 37 | Official leaderboard | |
| PinchBench 2.0 | 76% | #21 of 57 | Official leaderboard | |
| BrowseComp | 79.3% | #30 of 81 | Lab model card | |
| Chess Puzzles | 41% | #37 of 93 | Epoch AI run | |
| CEO-Bench | $70,620 | #8 of 19 | Official leaderboard | |
| MCP Atlas | 79.1% | #16 of 37 | Official leaderboard | |
| Mystery Game Puzzles | 28% | #37 of 84 | Epoch AI run | |
| DeepSearchQA | 89.1% | #11 of 23 | Lab model card | |
| Game Arena Bargaining | 188 | #19 of 39 | Official leaderboard | |
| STATE-Bench Main 0.7.0 | 53.4% | #2 of 4 | Official leaderboard | |
| Game Arena Checkers | 953 | #21 of 41 | Official leaderboard | |
| RuneBench | 4.70 | #39 of 76 | Official leaderboard | |
| AutomationBench | 13.4% | #24 of 42 | Official leaderboard | |
| Game Arena Word Association | 209 | #24 of 40 | Official leaderboard | |
| SeedClawBench | 64.1% | #3 of 5 | Official leaderboard | |
| Game Arena Word Art | 331 | #24 of 39 | Official leaderboard | |
| Game Arena Coin Game | 313 | #25 of 39 | Official leaderboard | |
| EBR-bench | 19.1% | #20 of 31 | Epoch AI run | |
| Game Arena Nine Men's Morris | 560 | #26 of 39 | Official leaderboard | |
| Game Arena Poker | -6.53 | #22 of 31 | Official leaderboard | |
| Kaggle Game Arena | 220 | #30 of 42 | Official leaderboard | |
| Terminal-Bench 2.1 Extended | 27.6% | #32 of 44 | Official leaderboard | |
| BrowseComp Extended | 14.2% | #31 of 42 | Official leaderboard | |
| CreativeWork | 28.3% | #3 of 4 | Official leaderboard | |
| Game Arena Go | 510 | #30 of 39 | Official leaderboard | |
| Game Arena Five in a Row | 353 | #16 of 20 | Official leaderboard | |
| Game Arena Four in a Row | 196 | #35 of 43 | Official leaderboard | |
| ApprenticeBench | 7.00% | #15 of 18 | Official leaderboard | |
| DRACO | 77.7% | #10 of 12 | Lab model card | |
| Game Arena Ultimate Tic-Tac-Toe | 64 | #33 of 39 | Official leaderboard | |
| Terminal-Bench 2.1 | 68.9% | #11 of 13 | Official leaderboard | |
| Game Arena Dots and Boxes | 570 | #35 of 40 | Official leaderboard | |
| Game Arena Dark Hex | 93 | #37 of 41 | Official leaderboard | |
| Game Arena Chess | 284 | #38 of 42 | Official leaderboard | |
| Clocktower Bench | 1,241 | #10 of 11 | Official leaderboard | |
| ARC-AGI-3 | 0.18% | #25 of 27 | Official leaderboard | |
| Game Arena Clobber | 31 | #37 of 39 | Official leaderboard | |
| Game Arena Reversi | 42 | #37 of 39 | Official leaderboard | |
| Game Arena Lines of Action | 0 | #39 of 39 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vals Finance Agent | 64.4% | #1 of 49 | Official leaderboard | |
| ProfBench | 61.3% | #2 of 72 | Official leaderboard | |
| Vals Index v1 | 72.2% | #1 of 34 | Official leaderboard | |
| MortgageTax | 70.3% | #3 of 87 | Official leaderboard | |
| OfficeQA | 86.3% | #1 of 14 | Lab model card | |
| MedCode | 54.9% | #7 of 96 | Official leaderboard | |
| CaseLaw v2 | 68.4% | #5 of 53 | Official leaderboard | |
| FORTE | 77.6% | #2 of 21 | Official leaderboard | |
| Workspace-Bench-Lite | 66.7% | #1 of 9 | Official leaderboard | |
| Vending-Bench 2 | $10,937 | #9 of 67 | Official leaderboard | |
| TaxEval v2 | 75.3% | #21 of 129 | Official leaderboard | |
| LegalBench | 85.3% | #24 of 138 | Official leaderboard | |
| GDPval-AA | 1,753 | #3 of 16 | Lab model card | |
| CorpFin v2 | 66.1% | #23 of 119 | Official leaderboard | |
| QF-Bench | 61.2% | #2 of 9 | Official leaderboard | |
| CommerceAgentBench | 52.3% | #3 of 13 | Official leaderboard | |
| YC-Bench | $1,714,511 | #14 of 59 | Official leaderboard | |
| CADGenBench | 36.9% | #3 of 11 | Official leaderboard | |
| CHI-Bench | 24.4% | #7 of 23 | Official leaderboard | |
| PhysicianBench | 29.3% | #4 of 13 | Official leaderboard | |
| HealthBench Professional | 47% | #2 of 6 | Official leaderboard | |
| Excel Modeling Benchmark | 63.7% | #27 of 76 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 6.67% | #30 of 80 | Official leaderboard | |
| Vals Index v2 | 56.1% | #26 of 65 | Official leaderboard | |
| Vending-Bench Arena | $5,838 | #13 of 32 | Official leaderboard | |
| Legal Research Bench | 38.5% | #41 of 93 | Official leaderboard | |
| E-Commerce Bench | 2.59 | #9 of 19 | Official leaderboard | |
| MedScribe | 83% | #46 of 97 | Official leaderboard | |
| Vals Finance Agent v2 | 51.5% | #40 of 80 | Official leaderboard | |
| JobBench | 44.5% | #8 of 16 | Official leaderboard | |
| GDPval Extended | 0.90% | #22 of 43 | Official leaderboard | |
| Agents' Last Exam | 21.1% | #17 of 31 | Official leaderboard | |
| APEX-Agents | 49.2% | #34 of 61 | Collected by Epoch AI, source not named | |
| TRLawBench | 9.02 | #19 of 34 | Official leaderboard | |
| Tax Agent Bench | 23.7% | #49 of 84 | Official leaderboard | |
| OfficeQA Pro | 46.6% | #7 of 12 | Official leaderboard | |
| APEX-1 | 62.2% | #42 of 71 | Official leaderboard | |
| Agent Startup Bench | 62.3% | #3 of 5 | Official leaderboard | |
| APEX-Agents 1.1 | 63.1% | #40 of 65 | Official leaderboard | |
| APEX-Accounting | 56.5% | #31 of 49 | Official leaderboard | |
| 3DCodeArena | 977 | #9 of 14 | Official leaderboard | |
| PolicyBench | 80.5% | #33 of 50 | Official leaderboard | |
| BigLaw Bench | 90.9% | #2 of 3 | Lab model card | |
| MedXpertQA MM Extended | 51% | #29 of 41 | Official leaderboard | |
| EnterpriseBench: CoreCraft | 35.9% | #37 of 51 | Official leaderboard | |
| prinzbench | 25.0 | #19 of 26 | Official leaderboard | |
| Tuesday Work Index | 38.0 | #35 of 44 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vision Arena | 1,300 | #2 of 136 | Official leaderboard | |
| CharXiv | 91% | #5 of 75 | Lab model card | |
| ZeroBench | 15% | #28 of 76 | Official leaderboard | |
| MineBench | 1,593 | #32 of 75 | Official leaderboard | |
| Drone-Bench | 68.5% | #9 of 19 | Official leaderboard | |
| ParseBench | 63.3% | #22 of 44 | Official leaderboard | |
| WorldBench | 53.7% | #4 of 8 | Official leaderboard | |
| ClockBench | 15% | #23 of 37 | Official leaderboard | |
| AgenticVBench | 22.1% | #6 of 9 | Official leaderboard | |
| Chartography | 16.5% | #29 of 43 | Official leaderboard | |
| BVB | 63.6 | #31 of 45 | Official leaderboard | |
| VoxelBench Image | 1,266 | #9 of 13 | Official leaderboard | |
| Furniture Assembly Benchmark | 33.3% | #25 of 35 | Official leaderboard | |
| Image2FloorPlan | 50.2 | #3 of 4 | Official leaderboard | |
| Blueprint-Bench 2 | 24.5% | #29 of 38 | Official leaderboard | |
| MMMU-Pro Extended | 28.3% | #30 of 38 | Official leaderboard | |
| VoxelBench | 1,393 | #47 of 56 | Official leaderboard | |
| CharXiv Extended | 76% | #32 of 38 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| nocot-bench (NCRI) | 103.7 | #15 of 185 | Official leaderboard | |
| World-Class Bio | 61.1% | #6 of 74 | Official leaderboard | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 67.9% | #7 of 71 | Official leaderboard | |
| Cybench | 96% | #2 of 18 | Lab model card | |
| BiomniBench | 63.9 | #1 of 8 | Official leaderboard | |
| BioTIER Refuse | 94.9% | #9 of 69 | Official leaderboard | |
| Gray Swan Agent Red Teaming (ART) benchmark | 0.10% | #2 of 14 | Lab model card | |
| Molecular Biology Capabilities Test | 56% | #18 of 90 | Official leaderboard | |
| Dictatorship Eval | 88.3% | #4 of 20 | Official leaderboard | |
| DNA Synthesis Screening Evasion | 80% | #1 of 5 | Lab model card | |
| UK AISI unprompted sabotage evaluation | 0% | #1 of 5 | Lab model card | |
| Malicious use of Claude Code (Anthropic) | 91.2% | #4 of 17 | Lab model card | |
| Higher-difficulty benign request evaluations (Anthropic) | 0.01% | #1 of 4 | Lab model card | |
| UK AISI sabotage continuation evaluation | 0% | #1 of 4 | Lab model card | |
| LAB-Bench FigQA | 78.6% | #3 of 11 | Lab model card | |
| Malicious computer use (Anthropic) | 89.3% | #4 of 14 | Lab model card | |
| KORA | 57.2% | #14 of 42 | Official leaderboard | |
| ExploitBench | 28% | #3 of 9 | Official leaderboard | |
| Minimal-LinuxBench (Anthropic) | 2.60% | #1 of 3 | Lab model card | |
| SpatialBench-Long | 29.8% | #11 of 27 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, computer use (Anthropic) | 0.46% | #7 of 14 | Lab model card | |
| EpiBench | 22.6% | #14 of 27 | Official leaderboard | |
| VariantBench | 35% | #14 of 27 | Official leaderboard | |
| MASK | 91.7% | #10 of 19 | Lab model card | |
| Single-turn harmful request evaluations (Anthropic) | 98% | #10 of 19 | Lab model card | |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 0.74% | #7 of 13 | Lab model card | |
| Impossible tasks (Anthropic) | 45% | #6 of 11 | Lab model card | |
| BioSecBench-Surveillance | 45.8% | #15 of 27 | Official leaderboard | |
| MetagenomicsBench | 35.3% | #15 of 26 | Official leaderboard | |
| TxBench-Antibody-Discovery | 36% | #18 of 30 | Official leaderboard | |
| Firefox 147 exploit evaluation (Anthropic) | 45.2% | #3 of 5 | Lab model card | |
| Virology Capabilities Test (VCT) | 50% | #19 of 31 | Lab model card | |
| Prompt injection robustness against adaptive attackers, coding (Anthropic) | 2.34% | #9 of 14 | Lab model card | |
| PCT (Polarized Contrastive Pairs) | 48.2% | #30 of 46 | Official leaderboard | |
| CyberGym | 73% | #23 of 34 | Lab model card | |
| TxBench-Oligo-Discovery | 31.9% | #21 of 30 | Official leaderboard | |
| TxBench-Preclinical-Pharmacology | 45.3% | #21 of 30 | Official leaderboard | |
| BBQ (Bias Benchmark for QA) | 81.3% | #18 of 25 | Lab model card | |
| CyberBench (Vals) | 52.1% | #19 of 26 | Official leaderboard | |
| BioSecBench-Function | 30.8% | #20 of 27 | Official leaderboard | |
| scBench-Long | 11.4% | #21 of 27 | Official leaderboard | |
| Single-turn benign request evaluations (Anthropic) | 0.28% | #15 of 19 | Lab model card | |
| Refusal to assist with AI safety R&D (Anthropic) | 33.1% | #4 of 5 | Lab model card | |
| SHADE-Arena | 1.50% | #14 of 16 | Lab model card | |
| ExploitGym | 1.38% | #16 of 18 | Official leaderboard | |
| BioTIER Permit | 82.8% | #62 of 69 | Official leaderboard | |
| Anthropic political even-handedness eval | 95.1% | #13 of 14 | Lab model card | |
| Over-eager behavior on GUI computer use tasks (Anthropic) | 21.1% | #5 of 5 | Lab model card | |
| Higher-difficulty harmful request evaluations (Anthropic) | 99% | #4 of 4 | Lab model card | |
| Self-preference bias (Anthropic) | 0.40 | #4 of 4 | Lab model card | |
| UK AISI cyber ranges | 0% | #4 of 4 | Lab model card | |
| Adherence to its constitution (Anthropic) | 0.74 | #3 of 3 | Lab model card | |
| Protocols (Anthropic life sciences) | 51.8% | #3 of 3 | Lab model card | |
| Destructive production evaluation (Anthropic) | 2.50% | #2 of 2 | Lab model card | |
| XBOW Validation Benchmarks | 98.5% | #1 of 1 | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SAGE (Vals) | 56.1% | #1 of 81 | Official leaderboard | |
| ACCoRD | 79.8 | #14 of 152 | Official leaderboard | |
| Conceptual Reasoning Index | 65.5 | #21 of 151 | Official leaderboard | |
| LMCA | 52.2% | #25 of 147 | Official leaderboard | |
| CAIS Risk Index | 37.8 | #12 of 67 | Official leaderboard | |
| Vals Multimodal Index | 67.4% | #10 of 33 | Official leaderboard | |
| CAIS Text Capabilities Index | 38.0 | #25 of 68 | Official leaderboard | |
| CAIS Vision Capabilities Index | 51.2 | #27 of 53 | Official leaderboard | |
| Anthropic ECI (AECI) | 155.1 | #6 of 6 | Lab model card |
Its lab's numbers put it at 159.6, 2.1 points above its ECI
The Self-Reported Index fits one score to the 24 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 157.3 to 161.8. How we compute it.
- GPQA Diamond 94.2
- ARC-AGI 92
- MMMLU 91.5
- CharXiv 91
- Global MMLU 89.9
- MILU 89.9
- DeepSearchQA 89.1
- SWE-bench Verified 87.6
- ScreenSpot-Pro 87.6
- INCLUDE 87
- OfficeQA 86.3
- OfficeQA Pro 80.6
- SWE-bench Multilingual 80.5
- BrowseComp 79.3
- OSWorld-Verified 78
- DRACO 77.7
- MCP Atlas 77.3
- ARC-AGI-2 75.8
- Terminal-Bench 2.0 69.4
- Vals Finance Agent 64.4
- SWE-Bench Pro 64.3
- Humanity's Last Exam 54.7
- SimpleQA Verified 44.9
- SWE-bench Multimodal 34.5
Its lab's numbers next to 12 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 |
|---|---|---|---|
| OfficeQA Pro | 80.6% OfficeQA Pro, exact-match grading (Table 8.1.A, 8.10.1) | 46.6% Leaderboard, Claude Agent SDK | +34 pts |
| Humanity's Last Exam | 54.7% Multidisciplinary reasoning Humanity's Last Exam (with tools) | 36.2% Leaderboard | +19 pts |
| Terminal-Bench 2.0 | 69.4% Agentic terminal coding Terminal-Bench 2.0 | 80.2% Leaderboard, WOZCODE | −11 pts |
| SimpleQA Verified | 44.9% Simple-QA-Verified net score (correct minus incorrect), closed book, printed on chart. Correct 44.9, unsure 38.7, incorrect 16.3 (6.3.3.1, Figures 6.3.3.1.A and B) | 51.7% Epoch AI run, xhigh | −6.8 pts |
| OSWorld-Verified | 78% Agentic computer use OSWorld-Verified | 83.6% Leaderboard, 100 steps, Pointer Agent | −5.6 pts |
| SWE-bench Verified | 87.6% Agentic coding SWE-bench Verified | 83.5% Epoch AI run, max | +4.1 pts |
| GPQA Diamond | 94.2% Graduate-level reasoning GPQA Diamond | 90.2% Epoch AI run, xhigh | +4.0 pts |
| MCP Atlas | 77.3% Scaled tool use MCP-Atlas | 79.1% Leaderboard, max | −1.8 pts |
| ARC-AGI | 92% ARC-AGI-1 | 93.5% Leaderboard, high | −1.5 pts |
| Vals Finance Agent | 64.4% Agentic financial analysis Finance Agent v1.1 | 64.4% Leaderboard | Same |
| Vending-Bench 2 | $10,937 Vending-Bench 2 final balance at max effort. $7,971 at high effort (8.10.4) | $10,937 Leaderboard | Same |
| ARC-AGI-2 | 75.8% ARC-AGI-2 | 75.8% Leaderboard, max | 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. 74 of them aren't in the launch post.
System Card: Claude Opus 4.7 Apr 16, 2026
Capabilities
- SWE-bench Verified 87.6
- SWE-Bench Pro 64.3
- SWE-bench Multilingual 80.5
- SWE-bench Multimodal 34.5
- Terminal-Bench 2.0 69.4
- BrowseComp 79.3
- MMMLU 91.5
- Humanity's Last Exam 46.9
- CharXiv 82.1
- OSWorld-Verified 78
- GPQA Diamond 94.2
- ScreenSpot-Pro 79.5
- OfficeQA 86.3
- OfficeQA Pro 80.6
- Vals Finance Agent 64.4
- MCP Atlas 77.3
- ARC-AGI 92
- ARC-AGI-2 75.8
- USAMO 2026 69.3
- GraphWalks
- OpenAI MRCR v2
- DeepSearchQA 89.1
- DRACO 77.7
- LAB-Bench FigQA 78.6
- Vending-Bench 2 10937
- GDPval-AA 1753
- Global MMLU 89.9
- MILU 89.9
- INCLUDE 87
- BioPipelineBench Verified (Anthropic) 83.6
- BioMysteryBench
- Structural Biology (Anthropic)
- Organic Chemistry (Anthropic) 77.2
- Phylogenetics (Anthropic) 79.6
- Protocols (Anthropic life sciences) 51.8
Hallucination
Safety
- Single-turn harmful request evaluations (Anthropic) 98
- Single-turn benign request evaluations (Anthropic) 0.28
- Higher-difficulty harmful request evaluations (Anthropic) 99
- Higher-difficulty benign request evaluations (Anthropic) 0.01
- Multi-turn testing (Anthropic)
- Child safety evaluations (Anthropic)
- Suicide and self-harm evaluations (Anthropic)
- Disordered eating evaluations (Anthropic)
- Election integrity evaluations (Anthropic)
- Malicious use of Claude Code (Anthropic) 91.2
- Malicious computer use (Anthropic) 89.3
Robustness
Alignment
- Impossible tasks (Anthropic) 45
- Agentic code behavior scores (Anthropic)
- Over-eager behavior on GUI computer use tasks (Anthropic) 21.1
- Anthropic automated behavioral audit
- Petri
- UK AISI unprompted sabotage evaluation 0
- UK AISI sabotage continuation evaluation 0
- Destructive production evaluation (Anthropic) 2.5
- Destructive actions (Anthropic)
- Adherence to its constitution (Anthropic) 0.74
- MASK 91.7
- Refusal to assist with AI safety R&D (Anthropic) 33.1
- Self-preference bias (Anthropic) 0.4
- DTBench