Claude Haiku 4.5 ranks first on 5 of the 187 benchmarks with five or more models scored.
Closed weights Small
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 | |
|---|---|---|---|---|
| Phare | 83.2% | #1 of 62 | Official leaderboard | |
| BullshitBench | 87.3% | #5 of 107 | Official leaderboard | |
| BullshitBench V2 | 77% | #15 of 129 | Official leaderboard | |
| WritingBench | 77.1 | #14 of 45 | Official leaderboard | |
| AraGen | 65.8% | #14 of 42 | Official leaderboard | |
| LMArena (Elo) | 1,414 | #115 of 335 | Official leaderboard | |
| Judgemark v2.1 | 79.7 | #16 of 42 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 65.0 | #55 of 119 | Official leaderboard | |
| HELM Arabic | 78.8 | #14 of 28 | Official leaderboard | |
| MMMLU | 83% | #38 of 71 | Lab model card | |
| Vectara Hallucination Leaderboard (HHEM) | 4.50% | #80 of 148 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 9.80% | #58 of 104 | Official leaderboard | |
| SEA-HELM | 64.9 | #41 of 71 | Official leaderboard | |
| HELM Arabic Enterprise | 63.6 | #14 of 23 | Official leaderboard | |
| StudyArena Text | 1,219 | #55 of 85 | Official leaderboard | |
| MultiChallenge | 50.5% | #30 of 42 | Official leaderboard | |
| DesignArena ASCII Art | 1,135 | #56 of 77 | Official leaderboard | |
| IFEval | 80.1% | #46 of 60 | Official leaderboard | |
| Visual Aesthetic Benchmark | 18.8% | #16 of 18 | Official leaderboard | |
| Extended NYT Connections | 14.3% | #94 of 103 | Official leaderboard | |
| SimpleQA Verified | 13.2% | #87 of 94 | Epoch AI run | |
| EQ-Bench 4 | 1,053 | #30 of 32 | Official leaderboard | |
| Antidote | 932 | #34 of 36 | Official leaderboard | |
| Last Translation Benchmark | 0.66 | #34 of 36 | Official leaderboard | |
| Document Arena | 1,419 | #37 of 39 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH Level 5 | 96.4% | #9 of 94 | Epoch AI run | |
| AIME 2025 | 96.3% | #18 of 143 | Lab model card | |
| GPQA Main | 60.5% | #25 of 60 | Official leaderboard | |
| A^3-Bench | 56.4% | #3 of 7 | Official leaderboard | |
| GPQA Diamond | 71.2% | #117 of 229 | Epoch AI run | |
| ForecastBench | 0.14 | #44 of 82 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 66.7% | #117 of 212 | Epoch AI run | |
| DTBench | 73.6% | #110 of 180 | Official leaderboard | |
| FrontierMath | 5.90% | #37 of 55 | Official leaderboard | |
| ARC-AGI | 47.7% | #69 of 95 | Official leaderboard | |
| ARC-AGI-2 | 4.03% | #73 of 98 | Official leaderboard | |
| FLTEval | 23.0 | #3 of 4 | Official leaderboard | |
| FrontierMath Tier 4 | 2.08% | #43 of 55 | Epoch AI run | |
| CORE-Bench | 11.1% | #16 of 17 | Official leaderboard | |
| BixBench3 | 1.50 | #19 of 20 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-bench Verified | 66.6% | #18 of 60 | Official leaderboard | |
| ProgramBench | 0% | #12 of 25 | Official leaderboard | |
| SWE-Bench Pro | 39.5% | #12 of 24 | Official leaderboard | |
| WeirdML | 45.4% | #62 of 112 | Official leaderboard | |
| Vibe Code Bench v1 | 11.7% | #11 of 17 | Official leaderboard | |
| ts-bench | 20% | #5 of 7 | Official leaderboard | |
| DesignArena Game Dev | 1,087 | #104 of 143 | Official leaderboard | |
| IOI (Vals) | 6.17% | #40 of 55 | Official leaderboard | |
| DesignArena Website | 1,115 | #110 of 151 | Official leaderboard | |
| DesignArena UI Component | 1,091 | #106 of 139 | Official leaderboard | |
| DesignArena Code | 1,105 | #110 of 143 | Official leaderboard | |
| DesignArena 3D Design | 1,055 | #107 of 136 | Official leaderboard | |
| WebDev Arena | 1,329 | #88 of 105 | arena.ai, via Epoch AI | |
| DesignArena Data Visualization | 1,122 | #99 of 118 | Official leaderboard | |
| ProgramBench almost solved | 0% | #21 of 25 | Official leaderboard | |
| Code Arena | 1,330 | #108 of 128 | Official leaderboard | |
| Code Migration | 7.55% | #66 of 78 | Official leaderboard | |
| SWE-bench Multilingual | 64.7% | #11 of 13 | Official leaderboard | |
| Vibe Code Bench | 11.4% | #102 of 120 | Official leaderboard | |
| OTelBench | 5.80% | #12 of 14 | Official leaderboard | |
| DesignArena SVG | 1,016 | #89 of 103 | Official leaderboard | |
| ALE-Bench | 1,006 | #109 of 126 | Official leaderboard | |
| SWE-bench Verified (hard subset) | 36.6% | #8 of 9 | Lab model card | |
| ProgramBench tests passed | 30% | #23 of 25 | Official leaderboard | |
| SWE-Bench Pro V2 Hard | 25.5% | #18 of 18 | Official leaderboard | |
| GameDevBench | 18.6% | #17 of 17 | Official leaderboard | |
| FDE-Gym | -0.33 | #12 of 12 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BFCL V4 | 68.7% | #6 of 65 | Official leaderboard | |
| Terminal-Bench | 35.5% | #5 of 18 | tbench.ai, via Epoch AI | |
| CompileBench | 86.7% | #7 of 20 | Official leaderboard | |
| PinchBench | 71.6% | #17 of 48 | Official leaderboard | |
| SnakeBench | 26 | #74 of 188 | Official leaderboard | |
| BALROG | 31.2% | #14 of 33 | Official leaderboard | |
| MCP-Universe | 26.4% | #13 of 30 | Official leaderboard | |
| LLM Chess | 127 | #59 of 123 | Official leaderboard | |
| ClawBench | 30.1% | #3 of 6 | Official leaderboard | |
| CEO-Bench | $59,625 | #10 of 19 | Official leaderboard | |
| PACT | 1,510 | #15 of 26 | Official leaderboard | |
| tau2-bench | 76.6% | #47 of 79 | Lab model card | |
| PinchBench 2.0 | 67.7% | #35 of 57 | Official leaderboard | |
| Terminal-Bench 2.0 | 35.5% | #28 of 43 | Official leaderboard | |
| DeepResearch Bench | 45.5% | #15 of 22 | evals.futuresearch.ai, via Epoch AI | |
| Deep Research Bench (FutureSearch) | 0.46 | #15 of 21 | Official leaderboard | |
| Toolathlon | 26.2% | #36 of 50 | Official leaderboard | |
| OpenThoughts-TBLite | 64.4% | #3 of 4 | Official leaderboard | |
| ITBench | 42.1% | #10 of 13 | Official leaderboard | |
| LisanBench | 137.0 | #103 of 133 | Official leaderboard | |
| GAIA | 23.6% | #51 of 64 | Official leaderboard | |
| Clocktower Radio | 1,390 | #20 of 23 | Official leaderboard | |
| BalatroBench | 5.07 | #14 of 16 | Official leaderboard | |
| OSWorld-Verified | 50.7% | #36 of 41 | Lab model card | |
| Game Arena Chess Openings | 102 | #10 of 11 | Official leaderboard | |
| Chess Puzzles | 8.00% | #88 of 93 | Epoch AI run | |
| RuneBench | 1.64 | #72 of 76 | Official leaderboard | |
| AutomationBench | 0.46% | #40 of 42 | Official leaderboard | |
| o11y-bench | 26.6% | #21 of 22 | Official leaderboard | |
| SkillsBench | 30.1% | #21 of 22 | Official leaderboard | |
| MCP Atlas | 40.2% | #37 of 37 | Official leaderboard | |
| TUA-Bench | 23.9% | #12 of 12 | Official leaderboard | |
| tau2-bench Verified | 53.4% | #11 of 11 | Official leaderboard | |
| DeepSearchQA | 12.8% | #10 of 10 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MedScribe | 85.2% | #30 of 97 | Official leaderboard | |
| Vals Finance Agent v1 | 43% | #16 of 47 | Official leaderboard | |
| LMBuild | 55.1% | #7 of 18 | Official leaderboard | |
| CorpFin v2 | 60.6% | #66 of 119 | Official leaderboard | |
| CaseLaw v2 | 56.5% | #30 of 53 | Official leaderboard | |
| LegalBench | 81.2% | #79 of 138 | Official leaderboard | |
| MortgageTax | 62.2% | #53 of 87 | Official leaderboard | |
| Vals Finance Agent | 46.9% | #32 of 49 | Official leaderboard | |
| FailureSensorIQ | 61.8% | #16 of 24 | Official leaderboard | |
| CorpBench Work | 75.6 | #12 of 17 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 0.83% | #59 of 80 | Official leaderboard | |
| TaxEval v2 | 67.5% | #97 of 129 | Official leaderboard | |
| EnterpriseClawBench | 51.3% | #7 of 9 | Official leaderboard | |
| Vals Index v1 | 47.2% | #27 of 34 | Official leaderboard | |
| Vending-Bench Arena | $1,701 | #26 of 32 | Official leaderboard | |
| Public Benefits Bench | 54.3% | #42 of 50 | Official leaderboard | |
| MedCode | 32.7% | #81 of 96 | Official leaderboard | |
| Vending-Bench 2 | $459 | #58 of 67 | Official leaderboard | |
| PatientAgentBench | 3.63 | #7 of 8 | Official leaderboard | |
| Legal Research Bench | 10.6% | #82 of 93 | Official leaderboard | |
| Vals Finance Agent v2 | 31% | #71 of 80 | Official leaderboard | |
| FoodTruck Bench | $166 | #53 of 59 | Official leaderboard | |
| ProfBench | 39.6% | #65 of 72 | Official leaderboard | |
| Vals Index v2 | 22.9% | #59 of 65 | Official leaderboard | |
| PolicyBench | 74.5% | #46 of 50 | Official leaderboard | |
| Excel Modeling Benchmark | 22.7% | #70 of 76 | Official leaderboard | |
| Tax Agent Bench | 5.86% | #79 of 84 | Official leaderboard | |
| CHI-Bench | 6.20% | #22 of 23 | Official leaderboard | |
| EEBench | 3.50% | #30 of 30 | Official leaderboard | |
| RCM-Bench-1 | 70.2% | #20 of 20 | Official leaderboard | |
| TaxCalcBench | 13.7% | #17 of 17 | Official leaderboard | |
| 3DCodeArena | 771 | #14 of 14 | Official leaderboard | |
| QF-Bench | 20.8% | #9 of 9 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GeoGuesser environment (FineEnvs) | 0.43 | #4 of 14 | Official leaderboard | |
| MADQA | 68.2% | #6 of 18 | Official leaderboard | |
| MMMU | 73.2% | #51 of 103 | Lab model card | |
| CaP-X (CaP-Bench) | 15.2% | #6 of 10 | Official leaderboard | |
| FutureOmni | 52% | #2 of 3 | Official leaderboard | |
| vCX Benchmark | 82.8% | #30 of 42 | Official leaderboard | |
| ParseBench | 51.2% | #32 of 44 | Official leaderboard | |
| Blueprint-Bench 2 | 0% | #34 of 38 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIR-Bench 2024 | 93.2% | #1 of 82 | Official leaderboard | |
| LiveSecBench | 91.4 | #1 of 34 | Official leaderboard | |
| KORA | 70.2% | #2 of 42 | Official leaderboard | |
| Lakera B3 (Backbone Breaker Benchmark) | 87.5% | #2 of 35 | Official leaderboard | |
| Impossible tasks (Anthropic) | 30% | #1 of 11 | Lab model card | |
| Single-turn harmful request evaluations (Anthropic) | 99.4% | #3 of 19 | Lab model card | |
| Agentic coding malicious requests (Anthropic) | 100% | #1 of 5 | Lab model card | |
| Single-turn benign request evaluations (Anthropic) | 0.02% | #5 of 19 | Lab model card | |
| HELM Safety | 95.5 | #21 of 75 | Official leaderboard | |
| nocot-bench (NCRI) | 74.2 | #80 of 185 | Official leaderboard | |
| BrowseSafe-Adapt | 78.8% | #3 of 6 | Official leaderboard | |
| LinuxArena | 14.6% | #1 of 2 | Official leaderboard | |
| MCP prompt injection evaluation (Anthropic) | 92.5% | #1 of 2 | Lab model card | |
| BioTIER Refuse | 42% | #37 of 69 | Official leaderboard | |
| SimpleSafetyTests | 98.8% | #41 of 75 | Official leaderboard | |
| Computer use prompt injection evaluation (Anthropic) | 72.2% | #3 of 5 | Lab model card | |
| Cybench | 46.9% | #12 of 18 | Lab model card | |
| LAB-Bench CloningScenarios | 45% | #8 of 12 | Lab model card | |
| BrowseSafe++ | 72.8% | #11 of 15 | Official leaderboard | |
| Reward-hack-prone coding tasks (Anthropic) | 6.00% | #6 of 8 | Lab model card | |
| Molecular Biology Capabilities Test | 42.5% | #68 of 90 | Official leaderboard | |
| Human Pathogen Capabilities Test | 41.6% | #66 of 86 | Official leaderboard | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 42.5% | #55 of 71 | Official leaderboard | |
| BioTIER Permit | 98.3% | #54 of 69 | Official leaderboard | |
| Anthropic Red Team | 96.9% | #60 of 75 | Official leaderboard | |
| Average in-eval reward hacking rate (Anthropic) | 15.5% | #4 of 5 | Lab model card | |
| LAB-Bench FigQA | 49% | #9 of 11 | Lab model card | |
| LAB-Bench SeqQA | 66% | #10 of 12 | Lab model card | |
| BinaryAudit | 10.5% | #17 of 20 | Official leaderboard | |
| LAB-Bench ProtocolQA | 69% | #14 of 16 | Lab model card | |
| BBQ (Bias Benchmark for QA) | 71.2% | #22 of 25 | Lab model card | |
| ExploitBench | 14% | #8 of 9 | Official leaderboard | |
| World-Class Bio | 23.4% | #68 of 74 | Official leaderboard | |
| Gray Swan Agent Red Teaming (ART) benchmark | 51.1% | #13 of 14 | Lab model card | |
| Virology Capabilities Test (VCT) | 32% | #29 of 31 | Lab model card | |
| DeepsecBench | 4.69 | #28 of 29 | Official leaderboard | |
| PCT (Polarized Contrastive Pairs) | 62.2% | #45 of 46 | Official leaderboard | |
| VCT-v2 | 20% | #25 of 25 | Official leaderboard | |
| Malicious use of Claude Code (Anthropic) | 69.4% | #17 of 17 | Lab model card | |
| EvalDetectBench | 0.76 | #10 of 10 | Official leaderboard | |
| CRSBench | 31.4% | #5 of 5 | Official leaderboard | |
| BrowseSafe-Bench | 81% | #4 of 4 | Official leaderboard | |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 2.40% | #4 of 4 | Official leaderboard | |
| Tool use prompt injection evaluation (Anthropic) | 93.4% | #2 of 2 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HELM Capabilities | 71.7 | #20 of 60 | Official leaderboard | |
| ACCoRD | 66.7 | #64 of 152 | Official leaderboard | |
| LMCA | 30.9% | #88 of 147 | Official leaderboard | |
| Conceptual Reasoning Index | 43.1 | #91 of 151 | Official leaderboard | |
| CAIS Risk Index | 54.3 | #41 of 67 | Official leaderboard | |
| SAGE (Vals) | 31.8% | #71 of 81 | Official leaderboard | |
| CAIS Vision Capabilities Index | 37.5 | #48 of 53 | Official leaderboard | |
| Vals Multimodal Index | 42.9% | #30 of 33 | Official leaderboard | |
| CAIS Text Capabilities Index | 12.4 | #62 of 68 | Official leaderboard |
The Self-Reported Index fits one score to the 9 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 138.2 to 143.6. How we compute it.
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 |
|---|---|---|---|
| SWE-bench Verified | 73.3% Agentic coding SWE-bench Verified | 66.6% Leaderboard, high, mini-SWE-agent | +6.7 pts |
| GPQA Diamond | 73% Graduate-level reasoning GPQA Diamond | 71.2% Epoch AI run, 32K | +1.8 pts |
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)
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 25 of them aren't in the launch post.