- Agentic terminal coding Terminal-Bench 2.0 65.4
- Agentic coding SWE-bench Verified 80.8
- Agentic computer use OSWorld 72.7
- Mean of Agentic tool use τ2-bench Retail / Agentic tool use τ2-bench Telecom 95.6
- Scaled tool use MCP Atlas 59.5
- Agentic search BrowseComp 84
- Multidisciplinary reasoning Humanity's Last Exam (with tools) 53
- Agentic financial analysis Finance Agent 60.7
- Office tasks GDPVal-AA Elo 1606
- Novel problem-solving ARC AGI 2 68.8
- Graduate-level reasoning GPQA Diamond 91.3
- Visual reasoning MMMU Pro (with tools) 77.3
- Multilingual Q&A MMMLU 91.1
- BigLaw Bench 90.2%, quoted from Harvey in the launch post ("40% perfect scores and 84% above 0.8") 90.2
Claude Opus 4.6
Claude Opus 4.6 ranks first on 41 of the 343 benchmarks with five or more models scored.
363
Benchmarks with a score
41
Ranked first, out of five or more models
103
In the top three
Anthropic Feb 2026
Claude Opus 4.6
Closed weights Flagship
156.3
#37 of 256
Epoch Capabilities Index
Self-Reported Index 155.6 #57 of 404
- Humanity's Last Exam 34.4% #11 /47
- SWE-bench Verified 78.7% #4 /32
- SWE-Bench Pro 51.9% #4 /24
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) | 1,503 | #6 of 335 | Official leaderboard | |
| MMLU-Pro | 89.1% | #4 of 212 | Official leaderboard | |
| Judgemark v4 | 90.7 | #1 of 40 | Official leaderboard | |
| BullshitBench | 92.7% | #3 of 107 | Official leaderboard | |
| BullshitBench V2 | 87% | #5 of 129 | Official leaderboard | |
| Judgemark v2.1 | 92.9 | #2 of 42 | Official leaderboard | |
| EQ-Bench 3 | 1,712 | #4 of 78 | Official leaderboard | |
| Document Arena | 1,497 | #3 of 39 | Official leaderboard | |
| KINA | 49.9% | #3 of 37 | Official leaderboard | |
| MMMLU | 91.1% | #6 of 71 | Lab model card | |
| Visual Aesthetic Benchmark | 35.5% | #2 of 18 | Official leaderboard | |
| OpenAI MRCR v2 | 71.8% | #6 of 47 | Official leaderboard | |
| Phare | 74.1% | #9 of 62 | Official leaderboard | |
| Creative Writing v3 | 78.3 | #20 of 120 | Official leaderboard | |
| HLE-Verified (text-only Full) | 46.8% | #1 of 6 | Official leaderboard | |
| Extended NYT Connections | 92.1% | #19 of 103 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 77.7 | #24 of 119 | Official leaderboard | |
| MultiNRC | 57.1% | #8 of 38 | Official leaderboard | |
| Humanity's Last Exam | 34.4% | #11 of 47 | Official leaderboard | |
| DesignArena ASCII Art | 1,250 | #19 of 77 | Official leaderboard | |
| CL-bench Life | 17% | #4 of 16 | Official leaderboard | |
| FACTS Search | 76.7% | #8 of 31 | Official leaderboard | |
| CL-bench | 20.7% | #7 of 24 | Official leaderboard | |
| HUMAINE | 30 | #15 of 50 | Official leaderboard | |
| Lech Mazur Creative Writing | 1.08 | #15 of 50 | Official leaderboard | |
| STALE | 28.3% | #3 of 10 | Official leaderboard | |
| SimpleQA Verified | 45.2% | #18 of 56 | Official leaderboard | |
| Hemingway-bench | 1,045 | #15 of 42 | Official leaderboard | |
| Antidote | 1,042 | #13 of 36 | Official leaderboard | |
| Structured Output Benchmark | 85.3% | #9 of 19 | Official leaderboard | |
| Honesty human feedback win rate (Anthropic) | 61% | #3 of 6 | Lab model card | |
| SkateBench v2 | 64.4% | #11 of 21 | Official leaderboard | |
| EQ-Bench 4 | 1,210 | #17 of 32 | Official leaderboard | |
| MultiChallenge | 56% | #24 of 42 | Official leaderboard | |
| 100Q-Hard (Anthropic) | 21.5% | #5 of 7 | Lab model card | |
| Vectara Hallucination Leaderboard (2025 dataset) | 12.2% | #83 of 104 | Official leaderboard | |
| False-premise questions (Anthropic) | 96.8% | #6 of 6 | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIME 2025 | 99.8% | #9 of 143 | Lab model card | |
| FrontierMath | 40.7% | #6 of 55 | Official leaderboard | |
| AstaBench | 55.3% | #2 of 13 | Official leaderboard | |
| HMMT 2026 | 96.2% | #5 of 31 | Official leaderboard | |
| FrontierMath Tier 4 | 22.9% | #9 of 55 | Epoch AI run | |
| MathArena (Apex, IMO, USAMO) | 34.5% | #8 of 47 | Official leaderboard | |
| Echo General AI Prediction Leaderboard | 1,008 | #3 of 17 | Official leaderboard | |
| MathArena Apex Shortlist | 86.7% | #7 of 38 | Official leaderboard | |
| GPQA Diamond | 90.5% | #43 of 229 | Epoch AI run | |
| ProofBench | 50% | #12 of 59 | Official leaderboard | |
| DTBench | 91.2% | #39 of 180 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 94.4% | #52 of 212 | Epoch AI run | |
| FLTEval | 39.6 | #1 of 4 | Official leaderboard | |
| SimpleBench | 67.6% | #27 of 106 | Official leaderboard | |
| AIME 2026 | 96.7% | #8 of 31 | Official leaderboard | |
| ResearchClawBench | 21.5 | #6 of 23 | Official leaderboard | |
| ARC-AGI | 94% | #25 of 95 | Official leaderboard | |
| ARC-AGI-2 | 69.2% | #28 of 98 | Official leaderboard | |
| BTF-2 | 0.13 | #1 of 3 | Official leaderboard | |
| SpatialBench (LatchBio) | 52.8% | #5 of 14 | Official leaderboard | |
| HorizonMath | 2.70% | #3 of 8 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 66% | #34 of 86 | Official leaderboard | |
| KellyBench | -11.20% | #2 of 5 | Official leaderboard | |
| NatureBench | 11.1% | #8 of 18 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 26.8% | #37 of 68 | Official leaderboard | |
| ForecastBench | 0.14 | #45 of 82 | Official leaderboard | |
| IMProofBench | 31.7% | #5 of 9 | Official leaderboard | |
| USAMO 2026 | 47% | #6 of 9 | Official leaderboard | |
| BioPipelineBench (Anthropic) | 53.1% | #2 of 3 | Lab model card | |
| Phylogenetics (Anthropic) | 61.3% | #2 of 3 | Lab model card | |
| Leipzig Benchmark | 14% | #14 of 19 | Official leaderboard | |
| LABBench2 | 51.9% | #4 of 5 | Official leaderboard | |
| Organic Chemistry (Anthropic) | 53.9% | #4 of 5 | Lab model card | |
| BioMysteryBench | 61.5% | #9 of 10 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveSQLBench | 38% | #2 of 36 | Official leaderboard | |
| METR Time Horizon | 12.0 h | #2 of 26 | Collected by Epoch AI, source not named | |
| METR Time Horizon 1.1 | 12.0 h | #2 of 26 | Official leaderboard | |
| Snorkel Agentic Coding | 65.2% | #1 of 11 | Official leaderboard | |
| AutoLab | 0.68 | #1 of 10 | Official leaderboard | |
| WeirdML | 78% | #12 of 112 | Official leaderboard | |
| Vision2Web S2-2026 | 43.5% | #1 of 9 | Official leaderboard | |
| Vibe Code Bench v1 | 36.1% | #2 of 17 | Official leaderboard | |
| OpenHands Index | 66.7% | #4 of 33 | Official leaderboard | |
| SWE-bench Verified | 78.7% | #4 of 32 | Epoch AI run | |
| Spider 2.0 | 86.3% | #3 of 23 | Official leaderboard | |
| Bird-SQL | 70.2% | #2 of 13 | Official leaderboard | |
| SWE-Bench Pro | 51.9% | #4 of 24 | Official leaderboard | |
| MLE-bench | 63.1% | #2 of 12 | Official leaderboard | |
| SWE-Bench Pro Private | 47.1% | #2 of 12 | Official leaderboard | |
| QwenWebDev | 1,617 | #1 of 6 | Official leaderboard | |
| SWE-fficiency | 15.5% | #3 of 17 | Official leaderboard | |
| RoadmapBench | 32.2% | #2 of 11 | Official leaderboard | |
| DesignArena Website | 1,284 | #28 of 151 | Official leaderboard | |
| FeatureBench | 69.5% | #3 of 16 | Official leaderboard | |
| FeatureBench Lite | 69.5% | #3 of 16 | Official leaderboard | |
| DataPrep-Bench General | 78.2 | #1 of 5 | Official leaderboard | |
| DataPrep-Bench Medical | 40.5 | #1 of 5 | Official leaderboard | |
| MiMo Coding Bench | 77.1% | #1 of 5 | Official leaderboard | |
| DesignArena Data Visualization | 1,277 | #24 of 118 | Official leaderboard | |
| SWE-bench Verified (hard subset) | 47.2% | #2 of 9 | Lab model card | |
| DesignArena Code | 1,279 | #32 of 143 | Official leaderboard | |
| EvoCode-Bench | 44% | #3 of 13 | Official leaderboard | |
| GSO-Bench | 41.2% | #7 of 30 | Official leaderboard | |
| DesignArena Game Dev | 1,268 | #35 of 143 | Official leaderboard | |
| DesignArena 3D Design | 1,257 | #34 of 136 | Official leaderboard | |
| SWE-bench-Live | 63% | #3 of 12 | Official leaderboard | |
| OJBench | 67.2% | #2 of 8 | Official leaderboard | |
| DesignArena UI Component | 1,275 | #36 of 139 | Official leaderboard | |
| WebDev Arena | 1,556 | #30 of 105 | WebDev Arena Leaderboard, via Epoch AI | |
| Text-based reinforcement learning task (Anthropic) | 0.97 | #2 of 7 | Lab model card | |
| DesignArena SVG | 1,218 | #31 of 103 | Official leaderboard | |
| SWE-bench Multilingual | 77.8% | #22 of 73 | Lab model card | |
| PostTrainBench | 24.8% | #7 of 23 | Official leaderboard | |
| Internal AI research evaluation suite 2 | 0.61 | #2 of 6 | Lab model card | |
| Code Arena | 1,547 | #46 of 128 | Official leaderboard | |
| SWE-Together | 62.4% | #8 of 22 | Official leaderboard | |
| DesignArena Full Stack | 1,199 | #18 of 49 | Official leaderboard | |
| DesignArena React Native | 1,202 | #22 of 56 | Official leaderboard | |
| DataPrep-Bench Finance | 55.5 | #2 of 5 | Official leaderboard | |
| ALE-Bench | 1,607 | #51 of 126 | Official leaderboard | |
| DesignArena Agentic Overall | 1,190 | #31 of 74 | Official leaderboard | |
| DesignArena Agentic Game Dev | 1,146 | #12 of 27 | Official leaderboard | |
| DesignArena Android (Kotlin) | 1,180 | #21 of 46 | Official leaderboard | |
| APEX-SWE | 41.8% | #31 of 65 | Official leaderboard | |
| NanoGPT-Bench | 9.30% | #1 of 2 | Official leaderboard | |
| OpenRCA | 34.9% | #1 of 2 | Lab model card | |
| Vibe Code Bench | 57.6% | #61 of 120 | Official leaderboard | |
| Image-to-WebDev Arena | 1,544 | #28 of 55 | Official leaderboard | |
| GBAEval | 44.1% | #11 of 21 | Official leaderboard | |
| FrontierSWE v1 | 48.9% | #9 of 17 | Official leaderboard | |
| SlopCodeBench | 20.9% | #12 of 22 | Official leaderboard | |
| DesignArena Frontend Web App | 1,187 | #29 of 53 | Official leaderboard | |
| TasteVal | 40.4× | #11 of 20 | Official leaderboard | |
| ProgramBench | 0% | #14 of 25 | Official leaderboard | |
| QwenSVG | 1,541 | #4 of 7 | Official leaderboard | |
| ts-bench | 40% | #4 of 7 | Official leaderboard | |
| FrogsGame post-training task | 1.00% | #10 of 17 | Official leaderboard | |
| SWE Atlas | 33.3% | #13 of 22 | Official leaderboard | |
| DataPrep-Bench Law | 75.6 | #3 of 5 | Official leaderboard | |
| SWE Atlas Test Writing | 36.7% | #14 of 22 | Official leaderboard | |
| SWE Atlas Refactoring | 35.6% | #11 of 17 | Official leaderboard | |
| MirrorCode | 33.3% | #26 of 39 | Epoch AI run | |
| AlgoTune | 1.47× | #12 of 18 | Official leaderboard | |
| ProgramBench tests passed | 52.1% | #17 of 25 | Official leaderboard | |
| ProgramBench almost solved | 2.50% | #18 of 25 | Official leaderboard | |
| Data Agent Benchmark | 61% | #14 of 18 | Official leaderboard | |
| DataPrep-Bench Math | 24.1 | #4 of 5 | Official leaderboard | |
| DataPrep-Bench Science | 23.1 | #4 of 5 | Official leaderboard | |
| Vision2Web S3-2026 | 57.6% | #6 of 7 | Official leaderboard | |
| FrontierCode 1.1 | 26.9% | #7 of 8 | Official leaderboard | |
| Doc2Harness | 28.3% | #17 of 19 | Official leaderboard | |
| Internal model use survey (Anthropic) | 0% | #5 of 5 | Lab model card | |
| QwenVision2Code | 1,518 | #5 of 5 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Claw-Eval Pass^3 | 70.4% | #1 of 24 | Official leaderboard | |
| DeepResearch Bench | 55.3% | #1 of 22 | evals.futuresearch.ai, via Epoch AI | |
| Deep Research Bench (FutureSearch) | 0.55 | #1 of 21 | Official leaderboard | |
| VitaBench 2.0 | 50.3% | #1 of 17 | Official leaderboard | |
| LisanBench | 14206.3 | #8 of 133 | Official leaderboard | |
| tau2-bench | 95.6% | #5 of 79 | Lab model card | |
| DeepPlanning | 58.9% | #1 of 15 | Official leaderboard | |
| AutoResearchBench | 9.39% | #1 of 13 | Official leaderboard | |
| Claw-Eval-Live | 66.7% | #1 of 12 | Official leaderboard | |
| EnterpriseOps-Gym | 45.9% | #1 of 12 | Official leaderboard | |
| TERMS-Bench | 0.69 | #1 of 11 | Official leaderboard | |
| LLM-WikiRace | 16% | #2 of 21 | Official leaderboard | |
| SnakeBench | 33 | #18 of 188 | Official leaderboard | |
| Terminal-Bench 2.0 | 76.4% | #5 of 43 | Official leaderboard | |
| WebArena | 68% | #1 of 8 | Lab model card | |
| Gaia2-CLI | 57% | #1 of 7 | Official leaderboard | |
| Pencil Puzzle Bench | 33.3% | #8 of 53 | Official leaderboard | |
| Claw-Eval | 80.4% | #4 of 24 | Official leaderboard | |
| QwenClawBench | 59.5% | #2 of 12 | Official leaderboard | |
| ClawBench | 61.4% | #1 of 6 | Official leaderboard | |
| iOSWorld | 51.9% | #1 of 6 | Official leaderboard | |
| MM-ClawBench | 75.4% | #1 of 6 | Official leaderboard | |
| Toloka Arena | 33% | #1 of 5 | Official leaderboard | |
| Toolathlon | 47.2% | #11 of 50 | Official leaderboard | |
| MyPCBench | 58.2% | #2 of 9 | Official leaderboard | |
| LMArena Search Arena | 1,222 | #9 of 37 | Official leaderboard | |
| PinchBench | 74.4% | #12 of 48 | Official leaderboard | |
| AppWorld Test-Normal | 91.1% | #3 of 12 | Official leaderboard | |
| CAR-bench | 60% | #3 of 12 | Official leaderboard | |
| DeepSearchQA | 91.3% | #6 of 23 | Lab model card | |
| BrowseComp | 84% | #22 of 81 | Lab model card | |
| Game Arena Word Association | 303 | #14 of 40 | Official leaderboard | |
| WildClawBench | 51.6% | #12 of 31 | Official leaderboard | |
| AppWorld Test-Challenge | 72.4% | #5 of 12 | Official leaderboard | |
| Clocktower Radio | 1,625 | #10 of 23 | Official leaderboard | |
| LLM Chess | 198 | #54 of 123 | Official leaderboard | |
| Game Arena Clobber | 320 | #18 of 39 | Official leaderboard | |
| ScienceBoard | 14.6% | #7 of 15 | Official leaderboard | |
| Mystery Game Puzzles | 25% | #40 of 84 | Epoch AI run | |
| OSWorld-Verified | 72.7% | #20 of 41 | Lab model card | |
| SkillsBench | 50.2% | #11 of 22 | Official leaderboard | |
| Game Arena Werewolf | -1 | #4 of 8 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 284 | #20 of 39 | Official leaderboard | |
| MCP Atlas | 76.8% | #19 of 37 | Official leaderboard | |
| Game Arena Lines of Action | 380 | #21 of 39 | Official leaderboard | |
| PinchBench 2.0 | 69.9% | #32 of 57 | Official leaderboard | |
| Game Arena Word Art | 453 | #22 of 39 | Official leaderboard | |
| Game Arena Dots and Boxes | 865 | #23 of 40 | Official leaderboard | |
| τ³-bench Banking | 27.3% | #15 of 25 | Official leaderboard | |
| Game Arena Checkers | 892 | #25 of 41 | Official leaderboard | |
| Game Arena Poker | -1.90 | #19 of 31 | Official leaderboard | |
| Game Arena Bargaining | 172 | #24 of 39 | Official leaderboard | |
| Game Arena Coin Game | 316 | #24 of 39 | Official leaderboard | |
| RuneBench | 4.42 | #47 of 76 | Official leaderboard | |
| Kaggle Game Arena | 264 | #26 of 42 | Official leaderboard | |
| Game Arena Go | 570 | #26 of 39 | Official leaderboard | |
| DuetBench | 998 | #6 of 9 | Official leaderboard | |
| Game Arena Reversi | 324 | #27 of 39 | Official leaderboard | |
| Chess Puzzles | 21% | #65 of 93 | Epoch AI run | |
| ARC-AGI-3 | 0.51% | #19 of 27 | Official leaderboard | |
| Game Arena Nine Men's Morris | 527 | #28 of 39 | Official leaderboard | |
| Clocktower Bench | 1,432 | #8 of 11 | Official leaderboard | |
| Game Arena Four in a Row | 211 | #32 of 43 | Official leaderboard | |
| MobileWorld | 33.4% | #14 of 18 | Official leaderboard | |
| EBR-bench | 12.7% | #25 of 31 | Epoch AI run | |
| Game Arena Five in a Row | 315 | #17 of 20 | Official leaderboard | |
| Game Arena Dark Hex | 108 | #35 of 41 | Official leaderboard | |
| ApprenticeBench | 5.00% | #17 of 18 | Official leaderboard | |
| Game Arena Chess | 50 | #41 of 42 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vals Finance Agent v1 | 60.6% | #1 of 47 | Official leaderboard | |
| TaxEval v2 | 76% | #8 of 129 | Official leaderboard | |
| ProfBench | 58.8% | #5 of 72 | Official leaderboard | |
| Vals Finance Agent | 60% | #5 of 49 | Official leaderboard | |
| CorpFin v2 | 67% | #14 of 119 | Official leaderboard | |
| Vals Index v1 | 66% | #4 of 34 | Official leaderboard | |
| SpreadsheetBench 2 | 34.9% | #1 of 8 | Official leaderboard | |
| CHI-Bench | 28% | #3 of 23 | Official leaderboard | |
| LegalBench | 85.3% | #22 of 138 | Official leaderboard | |
| MortgageTax | 68.5% | #14 of 87 | Official leaderboard | |
| PRBench | 52.3% | #8 of 48 | Official leaderboard | |
| GDPval-MM | 73.5% | #1 of 6 | Official leaderboard | |
| TaxCalcBench | 52.9% | #3 of 17 | Official leaderboard | |
| AlphaEval | 64.4% | #1 of 5 | Official leaderboard | |
| FoodTruck Bench | $49,519 | #12 of 59 | Official leaderboard | |
| OccuBench | 71.5% | #3 of 14 | Official leaderboard | |
| MedScribe | 86.7% | #21 of 97 | Official leaderboard | |
| PhysicianBench | 31.7% | #3 of 13 | Official leaderboard | |
| Vending-Bench 2 | $8,018 | #18 of 67 | Official leaderboard | |
| MedCode | 49.1% | #26 of 96 | Official leaderboard | |
| FORTE | 73.2% | #6 of 21 | Official leaderboard | |
| EnterpriseClawBench | 63.5% | #3 of 9 | Official leaderboard | |
| MBABench | 18.2 | #2 of 6 | Official leaderboard | |
| GDPval-AA | 1,606 | #6 of 16 | Lab model card | |
| CaseLaw v2 | 62.1% | #20 of 53 | Official leaderboard | |
| DesignArena PDF | 1,202 | #14 of 37 | Official leaderboard | |
| APEX-1 | 65.8% | #27 of 71 | Official leaderboard | |
| YC-Bench | $1,269,734 | #24 of 59 | Official leaderboard | |
| OfficeQA Pro | 51.1% | #5 of 12 | Official leaderboard | |
| GDPval Extended | 1.10% | #19 of 43 | Official leaderboard | |
| QF-Bench | 59.2% | #4 of 9 | Official leaderboard | |
| E-Commerce Bench | 2.58 | #10 of 19 | Official leaderboard | |
| TRLawBench | 9.03 | #18 of 34 | Official leaderboard | |
| Vending-Bench Arena | $5,062 | #17 of 32 | Official leaderboard | |
| DesignArena Spreadsheet | 1,122 | #19 of 35 | Official leaderboard | |
| Remote Labor Index | 4.17% | #12 of 19 | Official leaderboard | |
| CADGenBench | 30.9% | #7 of 11 | Official leaderboard | |
| RCM-Bench-1 | 87.8% | #13 of 20 | Official leaderboard | |
| Real-World Finance (Anthropic) | 64.1% | #2 of 3 | Lab model card | |
| APEX-Agents | 46.3% | #42 of 61 | Collected by Epoch AI, source not named | |
| APEX-Agents 1.1 | 61.8% | #45 of 65 | Official leaderboard | |
| EnterpriseBench: CoreCraft | 30.8% | #38 of 51 | Official leaderboard | |
| BankerToolBench v1.1 | 7.40% | #8 of 10 | Official leaderboard | |
| prinzbench | 19.0 | #24 of 26 | Official leaderboard | |
| BigLaw Bench | 90.2% | #3 of 3 | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vision Arena | 1,299 | #4 of 136 | Official leaderboard | |
| Claw-Eval-MM | 54.7% | #1 of 11 | Official leaderboard | |
| MATH-Vision (MathVision) | 84.6% | #14 of 105 | Official leaderboard | |
| AgentWorldBench | 57.8 | #2 of 13 | Official leaderboard | |
| VoxelBench Image | 1,467 | #4 of 13 | Official leaderboard | |
| MeasureBench Synthetic | 17% | #6 of 19 | Official leaderboard | |
| MMMU-Pro | 77.3% | #30 of 90 | Lab model card | |
| MeasureBench Real-world | 18.7% | #7 of 19 | Official leaderboard | |
| QwenWorldBench | 56.1% | #2 of 5 | Official leaderboard | |
| VISTA | 46.1% | #23 of 54 | Official leaderboard | |
| ZeroBench | 12% | #34 of 76 | Official leaderboard | |
| OCR Arena | 1,679 | #5 of 11 | Official leaderboard | |
| MineBench | 1,472 | #41 of 75 | Official leaderboard | |
| BVB | 64.1 | #30 of 45 | Official leaderboard | |
| DesignArena Code to Video | 1,110 | #20 of 29 | Official leaderboard | |
| ParseBench | 53.8% | #31 of 44 | Official leaderboard | |
| CC-OCR-Bench-V2 | 58.8% | #11 of 15 | Official leaderboard | |
| CharXiv | 68.5% | #57 of 75 | Lab model card | |
| VoxelBench | 1,419 | #44 of 56 | Official leaderboard | |
| Furniture Assembly Benchmark | 28.3% | #28 of 35 | Official leaderboard | |
| ClockBench | 10% | #30 of 37 | Official leaderboard | |
| Math Kangaroo 2025 (MathArena) | 72.3% | #18 of 22 | Official leaderboard | |
| BabyVision | 14.8% | #25 of 27 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Human Pathogen Capabilities Test | 80.7% | #1 of 86 | Official leaderboard | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 75.2% | #2 of 71 | Official leaderboard | |
| BinaryAudit | 49.1% | #1 of 20 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, coding (Anthropic) | 0% | #1 of 14 | Lab model card | |
| nocot-bench (NCRI) | 104.3 | #14 of 185 | Official leaderboard | |
| Lakera B3 (Backbone Breaker Benchmark) | 85.5% | #3 of 35 | Official leaderboard | |
| EvalDetectBench | 0.89 | #1 of 10 | Official leaderboard | |
| Single-turn harmful request evaluations (Anthropic) | 99.6% | #2 of 19 | Lab model card | |
| HelpBench | 84% | #1 of 9 | Official leaderboard | |
| BBQ (Bias Benchmark for QA) | 90.9% | #3 of 25 | Lab model card | |
| DTap Direct Misuse | 21.2% | #1 of 8 | Official leaderboard | |
| DTap Prompt Injection | 8.60% | #1 of 8 | Official leaderboard | |
| KORA | 65.8% | #6 of 42 | Official leaderboard | |
| Creative biology | 60.3% | #1 of 7 | Lab model card | |
| FORTRESS | 13% | #11 of 73 | Official leaderboard | |
| BioTIER Refuse | 94.4% | #11 of 69 | Official leaderboard | |
| Cybench | 93% | #3 of 18 | Lab model card | |
| Pwn CTF challenges (Anthropic) | 71.4% | #1 of 6 | Lab model card | |
| UK AISI and CAISI simulated enterprise network range | 15.6 | #1 of 6 | Official leaderboard | |
| Web CTF challenges (Anthropic) | 100% | #1 of 6 | Lab model card | |
| CyberGym | 91% | #6 of 34 | Official leaderboard | |
| Molecular Biology Capabilities Test | 56.7% | #16 of 90 | Official leaderboard | |
| CRSBench | 82.3% | #1 of 5 | Official leaderboard | |
| World-Class Bio | 56.5% | #15 of 74 | Official leaderboard | |
| SHADE-Arena | 18% | #4 of 16 | Lab model card | |
| Reward-hack-prone coding tasks (Anthropic) | 0% | #2 of 8 | Lab model card | |
| Crypto CTF challenges (Anthropic) | 88.9% | #2 of 6 | Lab model card | |
| Network CTF challenges (Anthropic) | 100% | #2 of 6 | Lab model card | |
| Rev CTF challenges (Anthropic) | 100% | #2 of 6 | Lab model card | |
| CTI-REALM | 63.7% | #1 of 3 | Official leaderboard | |
| Dictatorship Eval | 83.5% | #7 of 20 | Official leaderboard | |
| Anthropic political even-handedness eval | 98.2% | #5 of 14 | Lab model card | |
| BiomniBench | 63.2 | #3 of 8 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 0.29% | #5 of 13 | Lab model card | |
| LABBench2 ProtocolQA2 | 58.4% | #2 of 5 | Official leaderboard | |
| Over-eager behavior on GUI computer use tasks (Anthropic) | 11.4% | #2 of 5 | Lab model card | |
| Refusal to assist with AI safety R&D (Anthropic) | 14% | #2 of 5 | Lab model card | |
| Gray Swan Agent Red Teaming (ART) benchmark | 0.20% | #6 of 14 | Lab model card | |
| Malicious computer use (Anthropic) | 88.3% | #7 of 14 | Lab model card | |
| SEC-bench Pro | 25.7% | #3 of 6 | Official leaderboard | |
| SEC-bench Pro 260617 | 30.8% | #3 of 6 | Official leaderboard | |
| Higher-difficulty harmful request evaluations (Anthropic) | 99.2% | #2 of 4 | Lab model card | |
| SSH stress-testing evaluation (Anthropic) | 95% | #1 of 2 | Lab model card | |
| LAB-Bench FigQA | 58% | #6 of 11 | Lab model card | |
| CyberGym-E2E | 62.6% | #3 of 5 | Official leaderboard | |
| LABBench2 FigQA2 | 28.7% | #3 of 5 | Official leaderboard | |
| LABBench2 LitQA3 | 75.6% | #3 of 5 | Official leaderboard | |
| Virology Capabilities Test (VCT) | 48.3% | #21 of 31 | Lab model card | |
| Malicious use of Claude Code (Anthropic) | 83.2% | #12 of 17 | Lab model card | |
| Impossible tasks (Anthropic) | 50% | #8 of 11 | Lab model card | |
| Higher-difficulty benign request evaluations (Anthropic) | 0.04% | #3 of 4 | Lab model card | |
| ExploitGym | 1.84% | #14 of 18 | Official leaderboard | |
| PCT (Polarized Contrastive Pairs) | 51.7% | #36 of 46 | Official leaderboard | |
| Agentic coding malicious requests (Anthropic) | 99.3% | #4 of 5 | Lab model card | |
| HarnessAudit-Bench | 0.29 | #6 of 7 | Official leaderboard | |
| BioTIER Permit | 86.5% | #60 of 69 | Official leaderboard | |
| DrugDiscoveryBench | 27.7% | #19 of 20 | Official leaderboard | |
| Single-turn benign request evaluations (Anthropic) | 0.68% | #19 of 19 | Lab model card | |
| Prompt injection robustness against adaptive attackers, computer use (Anthropic) | 17.8% | #14 of 14 | Lab model card | |
| LABBench2 SeqQA2 | 44.3% | #5 of 5 | Official leaderboard | |
| ASL-4 virology uplift trial (Anthropic) | 52% | #3 of 3 | Lab model card | |
| Chrome Production Commit Scanning (Google) | 54% | #3 of 3 | Official leaderboard | |
| Forensics CTF challenges (Anthropic) | 100% | #2 of 2 | Lab model card | |
| Internal codebase sabotage propensity (Anthropic) | 0% | #2 of 2 | Lab model card | |
| LinuxArena | 24.7% | #2 of 2 | Official leaderboard | |
| Misc CTF challenges (Anthropic) | 100% | #2 of 2 | Lab model card | |
| Participation in junk science (Anthropic) | 1.10% | #2 of 2 | Lab model card | |
| Creative Biology Uplift Trial (Anthropic) | 16.4 | #1 of 1 | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ACCoRD | 80.4 | #12 of 152 | Official leaderboard | |
| LMCA | 55.8% | #16 of 147 | Official leaderboard | |
| Conceptual Reasoning Index | 66.6 | #17 of 151 | Official leaderboard | |
| SAGE (Vals) | 51.6% | #12 of 81 | Official leaderboard | |
| CAIS Risk Index | 43.2 | #22 of 67 | Official leaderboard | |
| CAIS Text Capabilities Index | 37.5 | #26 of 68 | Official leaderboard | |
| TutorBench | 53.7% | #8 of 20 | Official leaderboard | |
| CAIS Vision Capabilities Index | 47.7 | #31 of 53 | Official leaderboard |
Its lab's numbers put it at 155.6, in line with its ECI
The Self-Reported Index fits one score to the 19 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 153.1 to 158.0. How we compute it.
- AIME 2025 99.8
- tau2-bench 95.6
- ARC-AGI 94
- GPQA Diamond 91.3
- DeepSearchQA 91.3
- MMMLU 91.1
- BrowseComp 84
- SWE-bench Verified 80.8
- SWE-bench Multilingual 77.8
- MMMU-Pro 77.3
- OSWorld-Verified 72.7
- ARC-AGI-2 68.8
- CharXiv 68.5
- Terminal-Bench 2.0 65.4
- BioMysteryBench 61.5
- Vals Finance Agent 60.7
- MCP Atlas 59.5
- Humanity's Last Exam 53
- SWE-bench Verified (hard subset) 47.2
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 | What differs |
|---|---|---|---|---|
| CyberGym | 66.6% CyberGym targeted vulnerability reproduction, pass@1 over 1,507 tasks, no thinking with think tool (2.17) | 91% Leaderboard, dynamic, mneme | −24 pts | The lab used its think-tool evaluation, while the board reports the later mneme agent running in a cleaned vulnerable Docker environment. |
| Humanity's Last Exam | 53% Multidisciplinary reasoning Humanity's Last Exam (with tools) | 34.4% Leaderboard, max | +19 pts | |
| MCP Atlas | 59.5% Scaled tool use MCP Atlas | 76.8% Leaderboard, max | −17 pts | |
| Terminal-Bench 2.0 | 65.4% Agentic terminal coding Terminal-Bench 2.0 | 76.4% Leaderboard, Meta-Harness | −11 pts | |
| SWE-bench Multilingual | 77.8% SWE-bench Multilingual, adaptive thinking, max effort, average of 25 trials (2.4) | 72% Leaderboard, mini-SWE-agent | +5.8 pts | |
| SWE-bench Verified | 80.8% Agentic coding SWE-bench Verified | 75.6% Leaderboard, mini-SWE-agent | +5.2 pts | |
| GPQA Diamond | 91.3% Graduate-level reasoning GPQA Diamond | 90.5% Epoch AI run, 32k thinking | +0.8 pts | |
| Vals Finance Agent | 60.7% Agentic financial analysis Finance Agent | 60% Leaderboard, thinking | +0.7 pts | |
| ARC-AGI-2 | 68.8% Novel problem-solving ARC AGI 2 | 69.2% Leaderboard, 120k, high | −0.4 pts | |
| ARC-AGI | 94% ARC-AGI-1, 120k thinking tokens, high effort, ARC Prize private set, reported by ARC Prize Foundation (2.9, Figure 2.9) | 94% Leaderboard, 120k, high | Same | |
| MMMU-Pro | 77.3% Visual reasoning MMMU Pro (with tools) | 77.3% Leaderboard, with tools | Same | |
| Vending-Bench 2 | $8,018 Vending-Bench 2 final balance, effort high (2.15) | $8,018 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. 80 of them aren't in the launch post.
System Card: Claude Opus 4.6 Feb 5, 2026
Capabilities
- SWE-bench Verified 80.8
- Terminal-Bench 2.0 65.4
- tau2-bench
- MCP Atlas 59.5
- OSWorld-Verified 72.7
- ARC-AGI-2 68.8
- GPQA Diamond 91.3
- MMMU-Pro 73.9
- MMMLU 91.1
- SWE-bench Multilingual 77.8
- OpenRCA 34.9
- ARC-AGI 94
- GDPval-AA 1606
- AIME 2025 99.8
- Vals Finance Agent 60.7
- Real-World Finance (Anthropic) 64.1
- Vending-Bench 2 8017.6
- CyberGym 66.6
- OpenAI MRCR v2
- GraphWalks
- LAB-Bench FigQA 58
- CharXiv 68.5
- WebArena 68
- BrowseComp 83.7
- Humanity's Last Exam 40
- DeepSearchQA 91.3
- BioPipelineBench (Anthropic) 53.1
- BioMysteryBench 61.5
- Structural Biology (Anthropic)
- Organic Chemistry (Anthropic) 53.9
- Phylogenetics (Anthropic) 61.3
Hallucination
Safety
- Single-turn harmful request evaluations (Anthropic) 99.6
- Single-turn benign request evaluations (Anthropic) 0.68
- Higher-difficulty harmful request evaluations (Anthropic) 99.2
- Higher-difficulty benign request evaluations (Anthropic) 0.04
- Multi-turn testing (Anthropic)
- Child safety evaluations (Anthropic)
- Suicide and self-harm evaluations (Anthropic)
- SSH stress-testing evaluation (Anthropic) 95
- Agentic coding malicious requests (Anthropic) 99.3
- Malicious use of Claude Code (Anthropic) 83.2
- Malicious computer use (Anthropic) 88.3
Robustness
Alignment
- External pilot use failure modes (Anthropic)
- Reward-hack-prone coding tasks (Anthropic) 0
- Impossible tasks (Anthropic) 50
- Agentic code behavior scores (Anthropic)
- Over-eager behavior on GUI computer use tasks (Anthropic) 11.4
- Anthropic automated behavioral audit
- Petri
- UK AISI alignment and monitorability evaluations
- Refusal to assist with AI safety R&D (Anthropic) 14
- Self-preference bias (Anthropic)
- Internal codebase sabotage propensity (Anthropic) 0
- Participation in junk science (Anthropic) 1.1
- Deference to governments in local languages (Anthropic)
- SHADE-Arena 18
- Agentic misalignment suite (Anthropic)
Bio and chem risk
Cyber risk
AI R&D risk
- SWE-bench Verified (hard subset) 47.2
- Internal model use survey (Anthropic) 0
- Kernels task (Anthropic) 190
- Time series forecasting task (Anthropic)
- Text-based reinforcement learning task (Anthropic) 0.975
- LLM training task (Anthropic) 34
- Quadruped reinforcement learning task (Anthropic)
- Novel compiler task (Anthropic)
- Internal AI research evaluation suite 2 0.612