- Agentic coding SWE-bench Verified 80.9
- Agentic terminal coding Terminal-bench 2.0 (128k thinking budget) 59.3
- Mean of Agentic tool use τ2-bench Retail / Agentic tool use τ2-bench Telecom 93.6
- Scaled tool use MCP Atlas 62.3
- Computer use OSWorld 66.3
- Novel problem solving ARC-AGI-2 (Verified) 37.6
- Graduate-level reasoning GPQA Diamond 87
- Visual reasoning MMMU (validation) 80.7
- Multilingual Q&A MMMLU 90.8
Claude Opus 4.5
Claude Opus 4.5 ranks first on 27 of the 246 benchmarks with five or more models scored.
264
Benchmarks with a score
27
Ranked first, out of five or more models
61
In the top three
Anthropic Nov 2025
Claude Opus 4.5
Closed weights Flagship
151.9
#61 of 256
Epoch Capabilities Index
Self-Reported Index 148.9 #102 of 404
- Humanity's Last Exam 25.2% #16 /47
- SWE-bench Verified 76.7% #9 /32
- SWE-Bench Pro 45.9% #6 /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 | |
|---|---|---|---|---|
| BullshitBench V2 | 90% | #4 of 129 | Official leaderboard | |
| Phare | 82.4% | #2 of 62 | Official leaderboard | |
| BullshitBench | 90.9% | #4 of 107 | Official leaderboard | |
| MMLU-Pro | 87.3% | #13 of 212 | Official leaderboard | |
| MMMLU | 90.8% | #7 of 71 | Lab model card | |
| LMArena (Elo) | 1,474 | #36 of 335 | Official leaderboard | |
| AraGen | 80.3% | #5 of 42 | Official leaderboard | |
| False-premise questions (Anthropic) | 1.20% | #1 of 6 | Lab model card | |
| WritingBench | 79.8 | #8 of 45 | Official leaderboard | |
| CL-bench | 21.1% | #5 of 24 | Official leaderboard | |
| EQ-Bench 3 | 1,522 | #20 of 78 | Official leaderboard | |
| Creative Writing v3 | 75.0 | #31 of 120 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 73.1 | #32 of 119 | Official leaderboard | |
| Visual Aesthetic Benchmark | 34.3% | #5 of 18 | Official leaderboard | |
| MultiNRC | 48.6% | #12 of 38 | Official leaderboard | |
| LOCA-Bench | 14.7% | #3 of 9 | Official leaderboard | |
| Humanity's Last Exam | 25.2% | #16 of 47 | Official leaderboard | |
| Role-play Benchmark (MiniMax) | 76.6 | #3 of 8 | Official leaderboard | |
| FACTS Search | 73.3% | #12 of 31 | Official leaderboard | |
| MultiChallenge | 59% | #18 of 42 | Official leaderboard | |
| 100Q-Hard (Anthropic) | 45.9% | #3 of 7 | Lab model card | |
| DesignArena ASCII Art | 1,185 | #34 of 77 | Official leaderboard | |
| SimpleQA Verified | 39% | #25 of 56 | Official leaderboard | |
| Document Arena | 1,468 | #18 of 39 | Official leaderboard | |
| HLE-Verified (text-only Full) | 38.8% | #3 of 6 | Official leaderboard | |
| Hemingway-bench | 1,016 | #22 of 42 | Official leaderboard | |
| HUMAINE | 25 | #31 of 50 | Official leaderboard | |
| FACTS Parametric | 30.6% | #28 of 45 | Official leaderboard | |
| SkateBench v2 | 50.3% | #14 of 21 | Official leaderboard | |
| Extended NYT Connections | 52.5% | #70 of 103 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 10.9% | #72 of 104 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CORE-Bench | 77.8% | #1 of 17 | Official leaderboard | |
| ChemBench | 65.5% | #4 of 24 | Official leaderboard | |
| SciPredict | 23.1% | #2 of 12 | Official leaderboard | |
| ForecastBench | 0.13 | #20 of 82 | Official leaderboard | |
| AIME 2025 | 92.8% | #36 of 143 | Lab model card | |
| FrontierScience-Research | 17.5% | #2 of 7 | Official leaderboard | |
| DTBench | 89.9% | #52 of 180 | Official leaderboard | |
| SimpleBench | 62% | #35 of 106 | Official leaderboard | |
| GPQA Diamond | 86% | #76 of 229 | Epoch AI run | |
| ProofBench | 36% | #20 of 59 | Official leaderboard | |
| IMO-ProofBench | 23.8% | #8 of 20 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 86.1% | #85 of 212 | Epoch AI run | |
| FrontierScience-Olympiad | 71.3% | #3 of 7 | Official leaderboard | |
| FrontierMath | 20.7% | #25 of 55 | Official leaderboard | |
| ARC-AGI-2 | 37.6% | #47 of 98 | Official leaderboard | |
| ARC-AGI | 80% | #48 of 95 | Official leaderboard | |
| FrontierMath Tier 4 | 4.17% | #33 of 55 | Epoch AI run | |
| SpatialBench (LatchBio) | 42.8% | #10 of 14 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 34.4% | #63 of 86 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 4.88% | #59 of 68 | Official leaderboard | |
| LABBench2 | 47.1% | #5 of 5 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| METR Time Horizon | 4.1 h | #1 of 31 | Official leaderboard | |
| SWE-fficiency | 22.5% | #1 of 17 | Official leaderboard | |
| OTelBench | 29% | #1 of 14 | Official leaderboard | |
| SWE-bench | 52.6% | #1 of 9 | Official leaderboard | |
| Text-based reinforcement learning task (Anthropic) | 0.97 | #1 of 7 | Lab model card | |
| OctoCodingbench | 36.2% | #1 of 6 | Official leaderboard | |
| Snorkel Agentic Coding | 58% | #2 of 11 | Official leaderboard | |
| Internal model use survey (Anthropic) | 50% | #1 of 5 | Lab model card | |
| SWE-Review | 16.2% | #1 of 5 | Official leaderboard | |
| VIBE (MiniMax) | 90.7% | #1 of 5 | Official leaderboard | |
| Vision2Web S1-2026 | 52.7% | #1 of 5 | Official leaderboard | |
| WeirdML | 63.7% | #26 of 112 | Official leaderboard | |
| OpenHands Index | 60.6% | #8 of 33 | Official leaderboard | |
| SWE-Bench Pro | 45.9% | #6 of 24 | Official leaderboard | |
| FeatureBench | 67.2% | #4 of 16 | Official leaderboard | |
| FeatureBench Lite | 67.2% | #4 of 16 | Official leaderboard | |
| METR Time Horizon 1.1 | 4.9 h | #7 of 26 | Official leaderboard | |
| AlgoTune | 1.77× | #5 of 18 | Official leaderboard | |
| SWE-bench Verified | 76.7% | #9 of 32 | Epoch AI run | |
| SWE-bench Verified (hard subset) | 46.7% | #3 of 9 | Lab model card | |
| Vision2Web S2-2026 | 40% | #3 of 9 | Official leaderboard | |
| CC-Bench-V2 Backend | 26.9% | #1 of 3 | Official leaderboard | |
| IOI (Vals) | 23.6% | #19 of 55 | Official leaderboard | |
| SWE-bench Multilingual | 76.2% | #27 of 73 | Lab model card | |
| DesignArena Data Visualization | 1,241 | #44 of 118 | Official leaderboard | |
| DesignArena Game Dev | 1,219 | #57 of 143 | Official leaderboard | |
| GSO-Bench | 26.5% | #12 of 30 | Official leaderboard | |
| Vibe Code Bench v1 | 20.6% | #7 of 17 | Official leaderboard | |
| DesignArena Code | 1,232 | #59 of 143 | Official leaderboard | |
| DesignArena Website | 1,240 | #63 of 151 | Official leaderboard | |
| WebDev Arena | 1,495 | #45 of 105 | arena.ai, via Epoch AI | |
| DesignArena UI Component | 1,229 | #60 of 139 | Official leaderboard | |
| DesignArena 3D Design | 1,196 | #61 of 136 | Official leaderboard | |
| DesignArena React Native | 1,183 | #27 of 56 | Official leaderboard | |
| Code Arena | 1,493 | #62 of 128 | Official leaderboard | |
| DesignArena SVG | 1,171 | #51 of 103 | Official leaderboard | |
| ALE-Bench | 1,429 | #63 of 126 | Official leaderboard | |
| Internal AI research evaluation suite 2 | 0.60 | #3 of 6 | Lab model card | |
| CC-Bench-V2 Frontend | 75.7% | #1 of 2 | Official leaderboard | |
| CC-Bench-V2 Multi-Step Chained Tasks | 61.6% | #1 of 2 | Official leaderboard | |
| PostTrainBench | 17.3% | #13 of 23 | Official leaderboard | |
| DesignArena Android (Kotlin) | 1,162 | #27 of 46 | Official leaderboard | |
| DesignArena Agentic Overall | 1,149 | #45 of 74 | Official leaderboard | |
| DesignArena Full Stack | 1,135 | #30 of 49 | Official leaderboard | |
| CF-Div2-Stepfun | 72.2% | #5 of 8 | Official leaderboard | |
| SWE-Bench Pro Private | 23.4% | #8 of 12 | Official leaderboard | |
| CC-Bench-V2 Large Repo Exploration | 64.5% | #2 of 3 | Official leaderboard | |
| SlopCodeBench | 17.4% | #15 of 22 | Official leaderboard | |
| DesignArena Frontend Web App | 1,141 | #37 of 53 | Official leaderboard | |
| TasteVal | 11.8× | #14 of 20 | Official leaderboard | |
| Vibe Code Bench | 20.6% | #86 of 120 | Official leaderboard | |
| WeirdML v3 | 6.51% | #22 of 27 | Official leaderboard | |
| DAXBench v2 | 59.6% | #43 of 47 | Official leaderboard | |
| Vals RSI Index | 0.15 | #26 of 27 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BFCL V4 | 77.5% | #1 of 65 | Official leaderboard | |
| LLM-WikiRace | 18% | #1 of 21 | Official leaderboard | |
| Elimination Game | 6 | #4 of 56 | Official leaderboard | |
| tau2-bench | 93.6% | #8 of 79 | Lab model card | |
| DeepResearch Bench | 54.8% | #3 of 22 | evals.futuresearch.ai, via Epoch AI | |
| MCPMark | 42.3% | #4 of 28 | Official leaderboard | |
| Deep Research Bench (FutureSearch) | 0.55 | #3 of 21 | Official leaderboard | |
| LisanBench | 5513.0 | #22 of 133 | Official leaderboard | |
| tau2-bench Verified | 82% | #2 of 11 | Official leaderboard | |
| SnakeBench | 30 | #40 of 188 | Official leaderboard | |
| VITA-Bench | 28.5% | #4 of 18 | Official leaderboard | |
| AgencyBench | 47.7% | #2 of 9 | Official leaderboard | |
| ITBench | 52.1% | #3 of 13 | Official leaderboard | |
| Terminal-Bench 2.0 | 63.1% | #10 of 43 | Official leaderboard | |
| GAIA | 74.1% | #15 of 64 | Official leaderboard | |
| OSWorld-Verified | 76.3% | #8 of 33 | Official leaderboard | |
| BalatroBench | 13.9 | #4 of 16 | Official leaderboard | |
| EnterpriseOps-Gym | 39.4% | #3 of 12 | Official leaderboard | |
| BrowseComp-Plus | 67.6% | #1 of 4 | Lab model card | |
| DeepPlanning | 37% | #4 of 15 | Official leaderboard | |
| tau3-bench | 85.3% | #2 of 7 | Official leaderboard | |
| WorldTravel | 21.3% | #2 of 7 | Official leaderboard | |
| Toolathlon | 43.5% | #15 of 50 | Official leaderboard | |
| BALROG | 43.5% | #10 of 33 | Official leaderboard | |
| PinchBench | 72.5% | #16 of 48 | Official leaderboard | |
| Step Race | 3 | #20 of 59 | Official leaderboard | |
| LLM Chess | 446 | #43 of 123 | Official leaderboard | |
| Pencil Puzzle Bench | 6.70% | #20 of 53 | Official leaderboard | |
| ToolcallFormatIFBench | 84.2% | #2 of 5 | Official leaderboard | |
| RWSearch | 75.5% | #3 of 7 | Official leaderboard | |
| WebArena | 65.3% | #4 of 8 | Lab model card | |
| Mystery Game Puzzles | 22% | #44 of 84 | Epoch AI run | |
| QwenClawBench | 56.1% | #7 of 12 | Official leaderboard | |
| SkillsBench | 49% | #13 of 22 | Official leaderboard | |
| Game Arena Four in a Row | 302 | #27 of 43 | Official leaderboard | |
| Agent Poker Bench | 1,033 | #11 of 17 | Official leaderboard | |
| MCP Atlas | 69.8% | #25 of 37 | Official leaderboard | |
| EBR-bench | 14.3% | #22 of 31 | Epoch AI run | |
| RuneBench | 4.08 | #55 of 76 | Official leaderboard | |
| Game Arena Chess Openings | 252 | #8 of 11 | Official leaderboard | |
| LMArena Search Arena | 1,186 | #27 of 37 | Official leaderboard | |
| τ³-bench Banking | 24.7% | #19 of 25 | Official leaderboard | |
| DeepSearchQA | 24% | #8 of 10 | Official leaderboard | |
| Chess Puzzles | 14% | #79 of 93 | Epoch AI run | |
| Multi-agent search (Anthropic internal benchmark) | 74.8% | #1 of 1 | Lab model card | |
| OSWorld | 66.3% | #1 of 1 | Claim in Anthropic annoucement, via Epoch AI |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vals Finance Agent v1 | 55.2% | #4 of 47 | Official leaderboard | |
| PLawBench | 66.5% | #1 of 11 | Official leaderboard | |
| MortgageTax | 68.7% | #12 of 87 | Official leaderboard | |
| Vals Finance Agent | 58.8% | #8 of 49 | Official leaderboard | |
| Finch (FinWorkBench) | 43% | #1 of 6 | Official leaderboard | |
| TaxEval v2 | 74.9% | #27 of 129 | Official leaderboard | |
| LegalBench | 84.6% | #31 of 138 | Official leaderboard | |
| MedCode | 49.2% | #25 of 96 | Official leaderboard | |
| ProfBench | 54.2% | #20 of 72 | Official leaderboard | |
| MedScribe | 85.3% | #28 of 97 | Official leaderboard | |
| CorpFin v2 | 65.1% | #35 of 119 | Official leaderboard | |
| FailureSensorIQ | 66.7% | #8 of 24 | Official leaderboard | |
| GDPval-MM | 61.1% | #2 of 6 | Official leaderboard | |
| CaseLaw v2 | 62.6% | #18 of 53 | Official leaderboard | |
| Vals Index v1 | 54.6% | #14 of 34 | Official leaderboard | |
| OccuBench | 65.2% | #6 of 14 | Official leaderboard | |
| APEX-Agents | 18.4% | #4 of 8 | Official leaderboard | |
| TaxCalcBench | 36.3% | #9 of 17 | Official leaderboard | |
| Vending-Bench 2 | $4,967 | #37 of 67 | Official leaderboard | |
| PRBench | 44.2% | #28 of 48 | Official leaderboard | |
| Vending-Bench Arena | $4,143 | #19 of 32 | Official leaderboard | |
| SpreadsheetBench | 64.3% | #4 of 6 | Lab model card | |
| Remote Labor Index | 3.75% | #13 of 19 | Official leaderboard | |
| APEX-1 | 57.3% | #59 of 71 | Official leaderboard | |
| OfficeQA Pro | 38.4% | #10 of 12 | Official leaderboard | |
| prinzbench | 14.0 | #25 of 26 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMMU | 80.7% | #6 of 136 | Official leaderboard | |
| DISBench | 28.7% | #1 of 9 | Official leaderboard | |
| MMLongBench-Doc | 61.9% | #2 of 17 | Official leaderboard | |
| WorldVQA | 36.8% | #3 of 22 | Official leaderboard | |
| OCR Arena | 1,778 | #2 of 11 | Official leaderboard | |
| MATH-Vision (MathVision) | 77.1% | #20 of 105 | Official leaderboard | |
| CaP-X (CaP-Bench) | 23.8% | #2 of 10 | Official leaderboard | |
| MMMU-Pro | 73.9% | #12 of 57 | Official leaderboard | |
| MotionBench | 65.9% | #3 of 12 | Official leaderboard | |
| VISTA | 46.4% | #22 of 54 | Official leaderboard | |
| GeoBench | 75% | #20 of 48 | Official leaderboard | |
| VisFactor | 24.5% | #12 of 26 | Official leaderboard | |
| ChartMuseum | 60.7% | #10 of 21 | Official leaderboard | |
| ZeroBench | 9.00% | #41 of 76 | Official leaderboard | |
| Drone-Bench | 59.2% | #11 of 19 | Official leaderboard | |
| VoxelBench Image | 1,320 | #8 of 13 | Official leaderboard | |
| MineBench | 1,308 | #49 of 75 | Official leaderboard | |
| Furniture Assembly Benchmark | 28.3% | #27 of 35 | Official leaderboard | |
| VoxelBench | 1,357 | #48 of 56 | Official leaderboard | |
| BabyVision | 14.2% | #27 of 27 | Official leaderboard | |
| VPCT | 40% | #1 of 1 | Collected by Epoch AI, source not named |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Lakera B3 (Backbone Breaker Benchmark) | 89% | #1 of 35 | Official leaderboard | |
| BaxBench | 68.6% | #1 of 34 | Official leaderboard | |
| Single-turn harmful request evaluations (Anthropic) | 99.8% | #1 of 19 | Lab model card | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 71.5% | #4 of 71 | Official leaderboard | |
| LAB-Bench ProtocolQA | 90.7% | #1 of 16 | Lab model card | |
| FORTRESS | 9.63% | #5 of 73 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, computer use (Anthropic) | 0% | #1 of 14 | Lab model card | |
| nocot-bench (NCRI) | 102.7 | #18 of 185 | Official leaderboard | |
| Human Pathogen Capabilities Test | 67.9% | #9 of 86 | Official leaderboard | |
| Reward-hack-prone coding tasks (Anthropic) | 0% | #1 of 8 | Lab model card | |
| SCONE-bench | $548,400,000 | #1 of 7 | Official leaderboard | |
| LAB-Bench CloningScenarios | 75.8% | #2 of 12 | Lab model card | |
| Crypto CTF challenges (Anthropic) | 94.4% | #1 of 6 | Lab model card | |
| Network CTF challenges (Anthropic) | 100% | #1 of 6 | Lab model card | |
| Rev CTF challenges (Anthropic) | 100% | #1 of 6 | Lab model card | |
| BioTIER Refuse | 92.6% | #12 of 69 | Official leaderboard | |
| KORA | 60% | #11 of 42 | Official leaderboard | |
| BinaryAudit | 36.8% | #6 of 20 | Official leaderboard | |
| Molecular Biology Capabilities Test | 52.9% | #28 of 90 | Official leaderboard | |
| BBQ (Bias Benchmark for QA) | 88.7% | #8 of 25 | Lab model card | |
| Cybench | 82% | #6 of 18 | Lab model card | |
| UK AISI and CAISI simulated enterprise network range | 11.0 | #2 of 6 | Official leaderboard | |
| ASL-4 virology uplift trial (Anthropic) | 66.8% | #1 of 3 | Lab model card | |
| Malicious computer use (Anthropic) | 88.4% | #5 of 14 | Lab model card | |
| Prompt injection robustness against adaptive attackers, coding (Anthropic) | 0.30% | #5 of 14 | Lab model card | |
| Agentic coding malicious requests (Anthropic) | 100% | #2 of 5 | Lab model card | |
| Creative biology | 52.4% | #3 of 7 | Lab model card | |
| World-Class Bio | 47.9% | #34 of 74 | Official leaderboard | |
| LAB-Bench SeqQA | 85.2% | #6 of 12 | Lab model card | |
| Web CTF challenges (Anthropic) | 92.3% | #3 of 6 | Lab model card | |
| Forensics CTF challenges (Anthropic) | 100% | #1 of 2 | Lab model card | |
| InterCode-CTF | 93.6% | #1 of 2 | Official leaderboard | |
| Internal codebase sabotage propensity (Anthropic) | 0% | #1 of 2 | Lab model card | |
| Misc CTF challenges (Anthropic) | 100% | #1 of 2 | Lab model card | |
| CyberGym | 50.6% | #18 of 34 | Official leaderboard | |
| SHADE-Arena | 7.70% | #9 of 16 | Lab model card | |
| LABBench2 ProtocolQA2 | 51.2% | #3 of 5 | Official leaderboard | |
| LAB-Bench FigQA | 54.9% | #7 of 11 | Lab model card | |
| Gray Swan Agent Red Teaming (ART) benchmark | 4.70% | #9 of 14 | Lab model card | |
| Pwn CTF challenges (Anthropic) | 42.9% | #4 of 6 | Lab model card | |
| CTI-REALM | 62.4% | #2 of 3 | Official leaderboard | |
| Training distribution hack rate (Anthropic) | 1.00% | #2 of 3 | Lab model card | |
| PCT (Polarized Contrastive Pairs) | 49.4% | #31 of 46 | Official leaderboard | |
| ABC Bench Liquid-Handling Robot | 68% | #25 of 34 | Official leaderboard | |
| Single-turn benign request evaluations (Anthropic) | 0.23% | #14 of 19 | Lab model card | |
| Virology Capabilities Test (VCT) | 47.7% | #23 of 31 | Lab model card | |
| CyberGym-E2E (615 tasks) | 19.2% | #3 of 4 | Official leaderboard | |
| LABBench2 FigQA2 | 25.7% | #4 of 5 | Official leaderboard | |
| LABBench2 SeqQA2 | 45.5% | #4 of 5 | Official leaderboard | |
| BioTIER Permit | 94.3% | #58 of 69 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 8.00% | #11 of 13 | Lab model card | |
| Anthropic political even-handedness eval | 96% | #12 of 14 | Lab model card | |
| Malicious use of Claude Code (Anthropic) | 77.8% | #16 of 17 | Lab model card | |
| Impossible tasks (Anthropic) | 55% | #11 of 11 | Lab model card | |
| LABBench2 LitQA3 | 73.2% | #5 of 5 | Official leaderboard | |
| Bioinformatics evaluations (Anthropic) | 73.7% | #1 of 1 | Lab model card | |
| Sycophancy on user-provided prompts (Anthropic) | 10% | #1 of 1 | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SAGE (Vals) | 52.1% | #8 of 81 | Official leaderboard | |
| ACCoRD | 79.1 | #16 of 152 | Official leaderboard | |
| Conceptual Reasoning Index | 59.1 | #42 of 151 | Official leaderboard | |
| CAIS Risk Index | 41.6 | #19 of 67 | Official leaderboard | |
| LMCA | 44.5% | #45 of 147 | Official leaderboard | |
| CAIS Text Capabilities Index | 32.3 | #33 of 68 | Official leaderboard | |
| TutorBench | 51.2% | #14 of 20 | Official leaderboard | |
| CAIS Vision Capabilities Index | 45.4 | #38 of 53 | Official leaderboard |
Its lab's numbers put it at 148.9, 3.0 points below its ECI
The Self-Reported Index fits one score to the 17 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 147.3 to 150.5. How we compute it.
- tau2-bench 93.6
- AIME 2025 92.8
- MMMLU 90.8
- GPQA Diamond 87
- SWE-bench Verified 80.9
- MMMU 80.7
- ARC-AGI 80
- SWE-bench Multilingual 76.2
- OSWorld-Verified 66.3
- MCP Atlas 62.3
- Terminal-Bench 2.0 59.3
- Vals Finance Agent 55.2
- SWE-Bench Pro 51.6
- SWE-bench Verified (hard subset) 46.7
- SimpleQA Verified 44.1
- ARC-AGI-2 37.6
- Humanity's Last Exam 30.8
Its lab's numbers next to 16 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 | 66.3% Computer use OSWorld | 76.3% Leaderboard, 100 steps, VLAA-GUI | −10.0 pts |
| MCP Atlas | 62.3% Scaled tool use MCP Atlas | 69.8% Leaderboard, high | −7.5 pts |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 8.00% Claude for Chrome internal adaptive evaluation, attack success rate, model only without safeguards, 100 attempts per environment, lower is better (Figure 5.2.2.3.A; 7.6% with old safeguards, 1.4% with improved safeguards) | 1.40% Leaderboard, improved safeguards, Claude for Chrome | −6.6 pts |
| SWE-Bench Pro | 51.6% SWE-bench Pro, 64k thinking, avg of 5 trials (Table 2.4.A; 52.0% with no thinking) | 45.9% Leaderboard, SWE-agent | +5.7 pts |
| Humanity's Last Exam | 30.8% Humanity's Last Exam, reasoning only without tools (Figure 2.16.A; 43.2% with search, tools-only configuration) | 25.2% Leaderboard, thinking | +5.6 pts |
| SWE-bench Multilingual | 76.2% SWE-bench Multilingual, no thinking and 64k thinking both 76.20% (Table 2.4.A) | 70.7% Leaderboard, mini-SWE-agent | +5.5 pts |
| SimpleQA Verified | 44.1% Simple-QA-Verified correct rate, extended thinking (Figure 4.1.B; incorrect 35.4%, uncertain 20.5%) | 39% Leaderboard | +5.1 pts |
| Terminal-Bench 2.0 | 59.3% Agentic terminal coding Terminal-bench 2.0 (128k thinking budget) | 63.1% Leaderboard, Droid | −3.9 pts |
| Vals Finance Agent | 55.2% FinanceAgent (Vals AI), external Vals AI analysis, 64k thinking, 200k context, avg of 8 trials (2.13; internal testing 61.07% same settings, 61.03% with 1M context) | 58.8% Leaderboard, thinking | −3.6 pts |
| SWE-bench Verified | 80.9% Agentic coding SWE-bench Verified | 79.2% Leaderboard, Sonar Foundation Agent | +1.7 pts |
| GPQA Diamond | 87% Graduate-level reasoning GPQA Diamond | 86% Epoch AI run, 32k thinking | +1.0 pts |
| ARC-AGI-2 | 37.6% Novel problem solving ARC-AGI-2 (Verified) | 37.6% Leaderboard, thinking, 64k | Same |
| ARC-AGI | 80% ARC-AGI-1, semi-private set, 64k thinking, reported by ARC Prize Foundation (2.10, Figure 2.10.A) | 80% Leaderboard, thinking, 64k | Same |
| CyberGym | 50.6% CyberGym, pass@1 over 1,505 tasks, avg of 5 replicas, no thinking (2.14) | 50.6% Leaderboard, Anthropic Agent | Same |
| MMMU | 80.7% Visual reasoning MMMU (validation) | 80.7% Leaderboard | Same |
| Vending-Bench 2 | $4,967 Vending-Bench 2 final balance, effort high, 8,192 reasoning tokens per turn (2.11) | $4,967 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. 68 of them aren't in the launch post.
System Card: Claude Opus 4.5 Nov 24, 2025
Capabilities
- SWE-bench Verified 80.9
- Terminal-Bench 2.0 59.3
- tau2-bench
- MCP Atlas 62.3
- OSWorld-Verified 66.3
- ARC-AGI-2 37.6
- GPQA Diamond 87
- MMMU 80.7
- MMMLU 90.8
- SWE-Bench Pro 51.6
- SWE-bench Multilingual 76.2
- BrowseComp-Plus 67.6
- Multi-agent search (Anthropic internal benchmark) 74.8
- ARC-AGI 80
- Vending-Bench 2 4967.1
- Vals Finance Agent 55.2
- CyberGym 50.6
- SpreadsheetBench 64.3
- Humanity's Last Exam 30.8
- AIME 2025 92.8
- LAB-Bench FigQA 54.9
- WebArena 65.3
Hallucination
- 100Q-Hard (Anthropic) 45.9
- SimpleQA Verified 44.1
- AA-Omniscience 51.9
- False-premise questions (Anthropic) 1.2
Safety
Robustness
Alignment
- Anthropic automated behavioral audit
- Petri
- Sycophancy on user-provided prompts (Anthropic) 10
- Agentic misalignment suite (Anthropic)
- Self-preference bias (Anthropic)
- Internal codebase sabotage propensity (Anthropic) 0
- Reward-hack-prone coding tasks (Anthropic) 0
- Impossible tasks (Anthropic) 55
- Training distribution hack rate (Anthropic) 1
- SHADE-Arena 7.7
- Subversion Strategy Eval
- UK AISI alignment and monitorability evaluations
Bio and chem risk
- Long-form virology tasks
- Virology Capabilities Test (VCT) 47.7
- DNA Synthesis Screening Evasion
- LAB-Bench ProtocolQA 90.7
- LAB-Bench SeqQA 85.2
- LAB-Bench CloningScenarios 75.8
- Creative biology 52.4
- Short-horizon computational biology tasks
- Bioinformatics evaluations (Anthropic) 73.7
- ASL-4 virology uplift trial (Anthropic) 66.8
Cyber risk
AI R&D risk
- SWE-bench Verified (hard subset) 46.7
- Kernels task (Anthropic) 252.4
- Time series forecasting task (Anthropic)
- Text-based reinforcement learning task (Anthropic) 0.975
- LLM training task (Anthropic) 16.5
- Quadruped reinforcement learning task (Anthropic)
- Novel compiler task (Anthropic)
- Internal AI research evaluation suite 2 0.604
- Internal model use survey (Anthropic) 50