Claude Sonnet 4 ranks first on 2 of the 146 benchmarks it has scores on.
Closed weights
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.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMLU-Pro | 83.7% | #38 of 212 | Official leaderboard | |
| LLM Confabulation Leaderboard | 13.2% | #13 of 58 | Official leaderboard | |
| Slop Score | 27.3 | #4 of 17 | Official leaderboard | |
| Judgemark v2.1 | 82.0 | #11 of 42 | Official leaderboard | |
| SuperGPQA | 57% | #16 of 58 | Official leaderboard | |
| Lech Mazur Creative Writing | 8.14 | #14 of 44 | Epoch AI | |
| MMMLU | 86.5% | #23 of 62 | Model card | |
| LMArena (Elo) | 1,402 | #127 of 333 | Official leaderboard | |
| FACTS Grounding | 79.1% | #13 of 34 | Official leaderboard | |
| WritingBench | 74.1 | #19 of 45 | Official leaderboard | |
| RewardBench 2 | 71.2% | #18 of 42 | Official leaderboard | |
| EQ-Bench 3 | 1,304 | #35 of 78 | Official leaderboard | |
| ECLeKTic | 18.5% | #17 of 35 | Official leaderboard | |
| Global MMLU Lite | 90.6% | #21 of 43 | Official leaderboard | |
| Vectara Hallucination Leaderboard (HHEM) | 4.50% | #73 of 148 | Official leaderboard | |
| WideSearch | 3.60% | #3 of 6 | Official leaderboard | |
| FaithJudge | 18.2% | #19 of 37 | Official leaderboard | |
| BullshitBench | 25.4% | #56 of 107 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 58.4 | #68 of 119 | Official leaderboard | |
| BullshitBench V2 | 30% | #71 of 122 | Official leaderboard | |
| Creative Writing v3 | 65.4 | #70 of 120 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 10.3% | #61 of 104 | Official leaderboard | |
| MultiChallenge | 57.1% | #21 of 35 | Official leaderboard | |
| FACTS Search | 66.3% | #19 of 31 | Official leaderboard | |
| FACTS Parametric | 20.4% | #34 of 45 | Official leaderboard | |
| Humanity's Last Exam | 7.76% | #36 of 47 | Official leaderboard | |
| Fiction.LiveBench | 46.9% | #45 of 57 | Epoch AI | |
| MultiNRC | 18.4% | #30 of 38 | Official leaderboard | |
| FACTS Multimodal | 28.6% | #31 of 39 | Official leaderboard | |
| FACTS Benchmark Suite | 28.6% | #37 of 40 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Misguided Attention | 68.6% | #2 of 45 | Official leaderboard | |
| CORE-Bench | 46.7% | #4 of 17 | Official leaderboard | |
| MATH Level 5 | 84.4% | #23 of 94 | Epoch AI | |
| HiPhO | 58% | #14 of 35 | Official leaderboard | |
| GPQA Diamond | 79.2% | #90 of 219 | Epoch AI | |
| OTIS Mock AIME 2024-2025 | 71.1% | #97 of 202 | Epoch AI | |
| AIME 2025 | 85% | #73 of 143 | Model card | |
| DTBench | 61.8% | #97 of 178 | Official leaderboard | |
| ForecastBench | 63.1 | #47 of 86 | Official leaderboard | |
| SimpleBench | 45.5% | #62 of 100 | Official leaderboard | |
| PHYBench | 19.2% | #18 of 28 | Official leaderboard | |
| ARC-AGI-2 | 5.93% | #62 of 95 | Official leaderboard | |
| CritPt | 0.29% | #108 of 160 | Epoch AI | |
| FrontierMath | 4.14% | #50 of 70 | Epoch AI | |
| ARC-AGI | 40% | #69 of 92 | Official leaderboard | |
| FrontierMath Tier 4 | 0% | #50 of 55 | Epoch AI |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-bench Verified | 76.8% | #3 of 60 | Official leaderboard | |
| SWE-bench Lite | 60.3% | #1 of 15 | Official leaderboard | |
| AutoCodeBench | 51.1% | #3 of 34 | Official leaderboard | |
| ArtifactsBench | 57.3% | #4 of 20 | Official leaderboard | |
| SWE-bench Multimodal | 35.6% | #2 of 9 | Official leaderboard | |
| Aider Polyglot | 61.3% | #15 of 54 | Epoch AI | |
| METR Time Horizon | 74 min | #9 of 31 | Official leaderboard | |
| SWE-Bench Pro | 42.7% | #9 of 24 | Official leaderboard | |
| WeirdML | 46.1% | #56 of 112 | Official leaderboard | |
| CodeClash | 1,224 | #4 of 8 | Official leaderboard | |
| Time series forecasting task (Anthropic) | 5.80 | #6 of 9 | Model card | |
| METR data deduplication | 27.6% | #2 of 3 | Model card | |
| IOI (Vals) | 6.50% | #38 of 55 | Official leaderboard | |
| GSO-Bench | 4.90% | #21 of 30 | Official leaderboard | |
| LiveCodeBench v5 | 57.9% | #15 of 21 | Official leaderboard | |
| LiveCodeBench v6 | 55.9% | #15 of 21 | Official leaderboard | |
| Text-based reinforcement learning task (Anthropic) | 0.68 | #5 of 7 | Model card | |
| ALE-Bench | 655 | #87 of 121 | Epoch AI | |
| LiveSQLBench | 20% | #26 of 36 | Official leaderboard | |
| SWE-bench Verified (hard subset) | 36.7% | #7 of 9 | Model card | |
| Internal AI research evaluation suite 2 | 0.36 | #5 of 6 | Model card | |
| Quadruped reinforcement learning task (Anthropic) | 0.35 | #8 of 8 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| tau-bench | 70.3% | #2 of 23 | Model card | |
| CompileBench | 93.3% | #2 of 20 | Official leaderboard | |
| SnakeBench | 32 | #26 of 183 | Official leaderboard | |
| Meta Agents Research Environments (ARE) / Gaia2 | 37.8% | #2 of 13 | Official leaderboard | |
| AgentBench FC | 58.2% | #2 of 12 | Official leaderboard | |
| Galileo Agent Leaderboard v2 | 55% | #4 of 21 | Official leaderboard | |
| MCP-Universe | 32.9% | #6 of 30 | Official leaderboard | |
| PinchBench | 74% | #14 of 48 | Official leaderboard | |
| LisanBench | 2236.0 | #42 of 133 | Official leaderboard | |
| Elimination Game | 5 | #18 of 56 | Official leaderboard | |
| MCPMark | 28.4% | #10 of 28 | Official leaderboard | |
| Terminal-Bench | 41.3% | #8 of 22 | Model card | |
| Step Race | 3 | #23 of 59 | Official leaderboard | |
| LLM Chess | 115 | #59 of 124 | Official leaderboard | |
| GAIA | 55.5% | #29 of 57 | Official leaderboard | |
| DeepResearch Bench | 46.6% | #13 of 24 | Epoch AI | |
| OSWorld | 43.9% | #5 of 9 | Epoch AI | |
| OSWorld-Verified | 43.9% | #20 of 33 | Official leaderboard | |
| Toolathlon | 29.9% | #31 of 50 | Official leaderboard | |
| LiveResearchBench | 62.3 | #5 of 7 | Official leaderboard | |
| PinchBench 2.0 | 48.8% | #49 of 57 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vals Finance Agent v1 | 44.5% | #10 of 47 | Official leaderboard | |
| TheAgentCompany | 33.1% | #3 of 14 | Epoch AI | |
| StockBench | 2.20% | #4 of 13 | Official leaderboard | |
| LegalBench | 83% | #63 of 138 | Official leaderboard | |
| CorpFin v2 | 61.2% | #56 of 119 | Official leaderboard | |
| TaxEval v2 | 72% | #66 of 129 | Official leaderboard | |
| MortgageTax | 62.5% | #51 of 86 | Official leaderboard | |
| MedCode | 35% | #73 of 96 | Official leaderboard | |
| ProfBench | 42.5% | #55 of 72 | Official leaderboard | |
| MedScribe | 72.4% | #80 of 98 | Official leaderboard | |
| TaxCalcBench | 23% | #16 of 17 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMMU | 74.4% | #24 of 136 | Official leaderboard | |
| VisualPuzzles | 47.3% | #7 of 29 | Official leaderboard | |
| VISTA | 45.5% | #25 of 54 | Official leaderboard | |
| Vision Arena | 1,207 | #75 of 136 | Official leaderboard | |
| OCRBench v2 | 42.4% | #36 of 65 | Official leaderboard | |
| ZeroBench | 6.00% | #38 of 66 | Official leaderboard | |
| Math-VR | 28.1% | #15 of 26 | Official leaderboard | |
| MindCube | 44.8% | #3 of 5 | Epoch AI | |
| EnigmaEval | 3.12% | #28 of 41 | Epoch AI | |
| Blueprint-Bench | 0.27 | #7 of 8 | Official leaderboard | |
| VPCT | 34% | #25 of 28 | Epoch AI | |
| ClockBench | 6.70% | #34 of 37 | Official leaderboard | |
| GeoBench | 37% | #47 of 48 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIR-Bench 2024 | 88.3% | #4 of 82 | Official leaderboard | |
| HELM Safety | 98.1 | #4 of 75 | Official leaderboard | |
| BaxBench | 56.1% | #2 of 34 | Official leaderboard | |
| SimpleSafetyTests | 100% | #15 of 75 | Official leaderboard | |
| Average in-eval reward hacking rate (Anthropic) | 14.3% | #1 of 5 | Model card | |
| FORTRESS | 18.1% | #22 of 66 | Official leaderboard | |
| Single-turn harmful request evaluations (Anthropic) | 99% | #6 of 18 | Model card | |
| Anthropic Red Team | 99.2% | #31 of 75 | Official leaderboard | |
| BBQ (Bias Benchmark for QA) | 86.3% | #10 of 24 | Model card | |
| Malicious use of Claude Code (Anthropic) | 90% | #7 of 16 | Model card | |
| Impossible tasks (Anthropic) | 45% | #5 of 11 | Model card | |
| Cybench | 35% | #11 of 24 | Official leaderboard | |
| Reward-hack-prone coding tasks (Anthropic) | 4.00% | #5 of 8 | Model card | |
| Spiral-Bench v1.2 | 32.1 | #14 of 21 | Official leaderboard | |
| Single-turn benign request evaluations (Anthropic) | 0.23% | #12 of 18 | Model card | |
| BinaryAudit | 14% | #15 of 20 | Official leaderboard | |
| CyberGym | 22.6% | #25 of 33 | Official leaderboard | |
| Network CTF challenges (Anthropic) | 25% | #5 of 6 | Model card | |
| Rev CTF challenges (Anthropic) | 37.5% | #5 of 6 | Model card | |
| Spiral-Bench v1.0 | 41.1 | #12 of 13 | Official leaderboard | |
| StrongREJECT | 2.24% | #21 of 22 | Model card | |
| Creative biology | 37% | #7 of 7 | Model card | |
| Crypto CTF challenges (Anthropic) | 31.8% | #6 of 6 | Model card | |
| Pwn CTF challenges (Anthropic) | 33.3% | #6 of 6 | Model card | |
| Web CTF challenges (Anthropic) | 73.3% | #6 of 6 | Model card | |
| Computer use prompt injection evaluation (Anthropic) | 69% | #5 of 5 | Model card | |
| Bioweapons acquisition uplift trial (Anthropic) | 1.70 | #3 of 3 | Model card | |
| Bioweapons knowledge questions | 51.5% | #3 of 3 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HELM Capabilities | 76.6 | #11 of 60 | Official leaderboard | |
| CAIS Risk Index | 47.1 | #25 of 65 | Official leaderboard | |
| LMCA | 29% | #87 of 139 | Epoch AI | |
| Conceptual Reasoning Index | 41.1 | #96 of 149 | Official leaderboard | |
| ACCoRD | 57.0 | #98 of 150 | Official leaderboard | |
| SAGE (Vals) | 35% | #67 of 81 | Official leaderboard | |
| CAIS Text Capabilities Index | 17.1 | #55 of 66 | Official leaderboard | |
| CAIS Vision Capabilities Index | 26.4 | #53 of 53 | Official leaderboard |
Anthropic's models in color, against every other model on the Epoch Capabilities Index, by release date.
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 42 of them aren't in the launch post.