- Agentic terminal coding Frontier-Bench v0.1 43.3
- Knowledge work GDPval-AA v2 1861
- Novel problem-solving ARC-AGI-3 (high) 30.2
- Agentic search BrowseComp 90.8
- Multidisciplinary reasoning Humanity's Last Exam (with tools) 64.7
- Computer use OSWorld 2.0 70.6
- Agentic coding DeepSWE v1.1 68.8
- Agentic coding FrontierCode v1.1, Main 53.4
- Business workflows AutomationBench 26
- Legal Legal Agent Benchmark, Held-out 11.7
- Health HealthBench Professional 59.8
- Best of Biology BioMysteryBench (hard) / Biology BioMysteryBench (human solved) 90.1
- SWE-bench Pro 79.2
- SWE-bench Multilingual 89.5
- SWE-bench Multimodal 59.4
- AA-Briefcase 1720
- ARC-AGI-1 97.5
- ARC-AGI-2 90.4
Claude Opus 5
Claude Opus 5 ranks first on 41 of the 243 benchmarks it has scores on.
261
Benchmarks with a score
41
Ranked first, out of five or more models
78
In the top three
Anthropic Jul 2026
Claude Opus 5
Closed weights Flagship
164.5
#8 of 271
Epoch Capabilities Index
- DeepSWE 73.7% #3 /28
- Terminal-Bench 4.0 53.9% #6 /23
- τ³-bench Banking 48.7% #2 /25
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 | |
|---|---|---|---|---|
| Longform Creative Writing (EQ-Bench) | 86.3 | #1 of 119 | Official leaderboard | |
| Global MMLU Lite | 95.7% | #1 of 43 | Official leaderboard | |
| EQ-Bench 4 | 1,365 | #1 of 32 | Official leaderboard | |
| Humanity's Last Exam | 64.7% | #5 of 147 | Model card | |
| LMArena (Elo) | 1,492 | #12 of 333 | Official leaderboard | |
| INCLUDE | 89.8% | #1 of 25 | Model card | |
| Creative Writing v3 | 81.3 | #6 of 120 | Official leaderboard | |
| BullshitBench V2 | 80.2% | #10 of 122 | Official leaderboard | |
| Extended NYT Connections | 94.3% | #8 of 97 | Official leaderboard | |
| Continual Learning Exam | 47.7% | #1 of 10 | Official leaderboard | |
| Document Arena | 1,490 | #5 of 37 | Official leaderboard | |
| Judgemark v4 | 78.8 | #7 of 40 | Official leaderboard | |
| BullshitBench | 60% | #19 of 107 | Official leaderboard | |
| SimpleQA Verified | 63.3% | #11 of 56 | Official leaderboard | |
| Global MMLU | 92.5% | #3 of 11 | Model card | |
| FACTS Search V2 | 85.3% | #6 of 21 | Official leaderboard | |
| AA-Omniscience | 49% | #5 of 17 | Model card | |
| MILU | 92.1% | #3 of 10 | Model card | |
| FACTS Parametric | 48.8% | #20 of 45 | Official leaderboard | |
| HLE-Diamond | 40.5% | #10 of 18 | Official leaderboard | |
| FACTS Parametric V2 | 48.8% | #14 of 20 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DTBench | 96% | #3 of 178 | Official leaderboard | |
| ProofBench | 78% | #1 of 59 | Official leaderboard | |
| NatureBench | 26.7% | #1 of 18 | Official leaderboard | |
| ARC-AGI-2 | 90.4% | #6 of 95 | Official leaderboard | |
| GPQA Diamond | 93.9% | #15 of 219 | Epoch AI | |
| CritPt | 29.1% | #12 of 160 | Epoch AI | |
| ProofBench v1.1 | 99% | #4 of 51 | Official leaderboard | |
| SimpleBench | 80.6% | #9 of 100 | Official leaderboard | |
| ARC-AGI | 97.5% | #9 of 92 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 98.9% | #20 of 202 | Epoch AI | |
| BTF-3 | 0.12 | #1 of 10 | Epoch AI | |
| ExplorationBench | 89% | #1 of 10 | Official leaderboard | |
| BioMysteryBench | 90.1% | #1 of 8 | Model card | |
| FrontierMath Tiers 1-3 v2 | 85.6% | #13 of 89 | Epoch AI | |
| SpatialBench (LatchBio) | 72.5% | #1 of 6 | Model card | |
| Human Behavior Bench | 80.2 | #2 of 11 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 73.2% | #13 of 71 | Epoch AI | |
| Terminal-Bench-Science | 30% | #4 of 17 | Official leaderboard | |
| MysteryMechanism | 37.4% | #7 of 29 | Official leaderboard | |
| ArXivMath | 90.8% | #2 of 8 | Model card | |
| LatchBio Capabilities v1.0 | 50% | #7 of 27 | Official leaderboard | |
| FrontierPhysics | 10.1% | #6 of 19 | Official leaderboard | |
| scBench (LatchBio SingleCellBench) | 60.6% | #2 of 6 | Model card | |
| ProteinGym | 47.7% | #1 of 3 | Model card | |
| Leipzig Benchmark | 75% | #8 of 19 | Official leaderboard | |
| SciUniverse | 30.5% | #3 of 6 | Official leaderboard | |
| PaperBenchX | 7.53% | #6 of 10 | Official leaderboard | |
| BixBench3 | 40.6 | #13 of 20 | Official leaderboard | |
| Riemann-bench | 60% | #3 of 4 | Model card | |
| FrontierMath Erdős | 0% | #11 of 14 | Epoch AI | |
| ForecastBench | 66.7 | #85 of 86 | Official leaderboard | |
| IMO 2026 | 100% | #1 of 1 | Model card | |
| Organic Chemistry V2 (Anthropic) | 61.6% | #1 of 1 | Model card | |
| Protein Design (Anthropic) | 42.5% | #1 of 1 | Model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-bench Verified | 96% | #1 of 150 | Model card | |
| IOI (Vals) | 91.7% | #1 of 55 | Official leaderboard | |
| AndroidBench | 91.8% | #1 of 32 | Official leaderboard | |
| SWE-bench Multilingual | 89.5% | #3 of 70 | Model card | |
| WeirdML | 91.8% | #5 of 112 | Official leaderboard | |
| SWE Atlas | 63.2% | #1 of 22 | Official leaderboard | |
| GBAEval | 79.6% | #1 of 21 | Official leaderboard | |
| ALE-Bench | 2,165 | #6 of 121 | Epoch AI | |
| SWE-Bench Pro V2 | 99.4% | #1 of 16 | Official leaderboard | |
| SWE-Bench Pro V2 Hard | 98% | #1 of 16 | Official leaderboard | |
| SWE-Bench Pro | 79.2% | #5 of 77 | Model card | |
| FrontierCode | 53.4% | #3 of 46 | Epoch AI | |
| WebDev Arena | 1,687 | #7 of 105 | Epoch AI | |
| Code Arena | 1,691 | #9 of 124 | Official leaderboard | |
| Vibe Code Bench | 88.4% | #10 of 115 | Official leaderboard | |
| SWE Atlas Test Writing | 62.2% | #2 of 22 | Official leaderboard | |
| PostTrainBench 1.1 | 45.6% | #1 of 11 | Official leaderboard | |
| Code Migration | 57.5% | #8 of 78 | Official leaderboard | |
| DeepSWE | 73.7% | #3 of 28 | Official leaderboard | |
| MirrorCode | 87.8% | #3 of 28 | Epoch AI | |
| Image-to-WebDev Arena | 1,662 | #6 of 55 | Official leaderboard | |
| Vibe Code Bench 1-100 | 28.5% | #3 of 26 | Official leaderboard | |
| Vals RSI Index | 0.33 | #3 of 25 | Official leaderboard | |
| Quadruped reinforcement learning task (Anthropic) | 31.3 | #1 of 8 | Model card | |
| SWE Refactor Bench | 32.3% | #1 of 8 | Official leaderboard | |
| IOI v2 (Vals) | 84.3% | #8 of 45 | Official leaderboard | |
| TasteVal | 128.6× | #4 of 20 | Official leaderboard | |
| SWE-bench Multimodal | 59.4% | #2 of 10 | Model card | |
| Vision2Web | 80.2% | #2 of 10 | Official leaderboard | |
| Senior SWE-Bench | 34.7% | #5 of 23 | Official leaderboard | |
| SWE-rebench | 63.4% | #3 of 13 | Official leaderboard | |
| PostTrainBench 1.2 | 41.7% | #4 of 17 | Official leaderboard | |
| ProgramBench | 4.50% | #6 of 25 | Official leaderboard | |
| Novel compiler task (Anthropic) | 80.9% | #1 of 4 | Model card | |
| MLS-Bench | 49.8% | #4 of 15 | Official leaderboard | |
| PostTrainBench | 35% | #3 of 11 | Epoch AI | |
| RSI-Exam | 46.4% | #5 of 18 | Official leaderboard | |
| FrontierSWE | 52% | #6 of 21 | Epoch AI | |
| ScarfBench | 69.8% | #3 of 10 | Official leaderboard | |
| FrontierCode 1.1 | 53.4% | #4 of 12 | Model card | |
| AutoResearchExam | 0.58 | #3 of 9 | Official leaderboard | |
| WeirdML v3 | 19% | #9 of 25 | Official leaderboard | |
| Brokk Power Ranking | 9.80 | #8 of 19 | Official leaderboard | |
| CursorBench | 46.6% | #8 of 19 | Epoch AI | |
| FrontierCode 2.0 | 26.7% | #11 of 26 | Official leaderboard | |
| Time series forecasting task (Anthropic) | 5.68 | #4 of 9 | Model card | |
| SWE-sweep | 1.30% | #7 of 10 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Game Arena Nine Men's Morris | 1,022 | #1 of 36 | Official leaderboard | |
| LisanBench | 15428.3 | #5 of 133 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 591 | #2 of 36 | Official leaderboard | |
| BALROG | 63.4% | #2 of 35 | Official leaderboard | |
| Kaggle Game Arena | 436 | #3 of 39 | Official leaderboard | |
| Toolathlon-Verified | 80.6% | #1 of 13 | Model card | |
| τ³-bench Banking | 48.7% | #2 of 25 | Official leaderboard | |
| Mystery Game Puzzles | 59% | #7 of 80 | Epoch AI | |
| MCP Atlas | 85.8% | #3 of 34 | Official leaderboard | |
| BrowseComp | 90.8% | #7 of 77 | Model card | |
| OSWorld-Verified | 83.4% | #3 of 33 | Official leaderboard | |
| DeepSearchQA | 95% | #2 of 22 | Model card | |
| DRACO | 88.6% | #1 of 11 | Model card | |
| Terminal-Bench 3.0 | 42.7% | #1 of 11 | Official leaderboard | |
| LLM Chess | 1,285 | #13 of 124 | Official leaderboard | |
| Game Arena Checkers | 1,173 | #4 of 38 | Official leaderboard | |
| Game Arena Four in a Row | 672 | #4 of 38 | Official leaderboard | |
| Game Arena Clobber | 566 | #4 of 36 | Official leaderboard | |
| Game Arena Go | 1,560 | #4 of 36 | Official leaderboard | |
| Game Arena Lines of Action | 938 | #4 of 36 | Official leaderboard | |
| Game Arena Reversi | 876 | #4 of 36 | Official leaderboard | |
| OSWorld 2.0 | 31.4% | #1 of 9 | Epoch AI | |
| Long-Horizon-Terminal-Bench (LHTB) | 51% | #3 of 25 | Official leaderboard | |
| Game Arena Dark Hex | 368 | #6 of 38 | Official leaderboard | |
| Game Arena Dots and Boxes | 1,077 | #6 of 37 | Official leaderboard | |
| Game Arena Poker | 10.0 | #6 of 31 | Official leaderboard | |
| RecreationBench | 44.2% | #2 of 10 | Official leaderboard | |
| ApprenticeBench | 36% | #4 of 18 | Official leaderboard | |
| Agent Arena | 0.08 | #12 of 53 | Official leaderboard | |
| ARC-AGI-3 | 30.2% | #6 of 24 | Official leaderboard | |
| SaaS-Bench | 31.1% | #1 of 4 | Official leaderboard | |
| Terminal-Bench 4.0 | 53.9% | #6 of 23 | Official leaderboard | |
| LMArena Search Arena | 1,214 | #10 of 35 | Official leaderboard | |
| EBR-bench | 45.7% | #8 of 28 | Epoch AI | |
| Game Arena Word Art | 797 | #11 of 35 | Official leaderboard | |
| Game Arena Coin Game | 401 | #13 of 35 | Official leaderboard | |
| Chess Puzzles | 43% | #31 of 83 | Epoch AI | |
| CUA-bench | 9.00% | #4 of 9 | Official leaderboard | |
| Time Horizon Index: KSP | 18.8% | #6 of 13 | Official leaderboard | |
| Brood War Bench | 66.7% | #3 of 6 | Official leaderboard | |
| OSWorld 2.1 | 44.3% | #2 of 4 | Official leaderboard | |
| Game Arena Word Association | 177 | #22 of 36 | Official leaderboard | |
| Game Arena Chess | 690 | #24 of 39 | Official leaderboard | |
| OSWorld 2.0 (Aug 2026) | 31.4% | #2 of 3 | Official leaderboard | |
| Game Arena Bargaining | 158 | #27 of 36 | Official leaderboard | |
| AutomationBench | 26% | #31 of 41 | Model card | |
| OSWorld 2.1 (offline subset) | 48.6% | #1 of 1 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CorpFin v2 | 73.2% | #1 of 119 | Official leaderboard | |
| MedCode | 63.6% | #1 of 96 | Official leaderboard | |
| MortgageTax | 72.1% | #1 of 86 | Official leaderboard | |
| ProfBench | 66.2% | #1 of 72 | Official leaderboard | |
| Public Benefits Bench | 76.9% | #1 of 50 | Official leaderboard | |
| Legal Research Bench | 55.3% | #2 of 88 | Official leaderboard | |
| GDPval-AA v2 | 1,861 | #1 of 26 | Model card | |
| MedScribe | 91% | #4 of 98 | Official leaderboard | |
| RCM-Bench-1 | 92.7% | #1 of 20 | Official leaderboard | |
| Tax Agent Bench | 45.5% | #5 of 79 | Official leaderboard | |
| LegalBench | 87% | #10 of 138 | Official leaderboard | |
| CommerceAgentBench | 61.7% | #1 of 13 | Official leaderboard | |
| Excel Modeling Benchmark | 73.6% | #6 of 76 | Official leaderboard | |
| Vending-Bench 2 | $11,182 | #6 of 66 | Official leaderboard | |
| FoodTruck Bench | $75,264 | #1 of 10 | Official leaderboard | |
| Vals Finance Agent v2 | 58.6% | #9 of 80 | Official leaderboard | |
| BlenderWorld | 73.0 | #2 of 13 | Official leaderboard | |
| APEX-Agents | 65.8% | #9 of 56 | Epoch AI | |
| AA-Briefcase | 1,720 | #1 of 6 | Model card | |
| TaxEval v2 | 75.1% | #23 of 129 | Official leaderboard | |
| Physical-Engineering's Last Exam | 0.43 | #4 of 18 | Official leaderboard | |
| OfficeQA | 78.1% | #4 of 12 | Model card | |
| Vending-Bench Arena | $7,044 | #11 of 32 | Official leaderboard | |
| EEBench | 61.6% | #9 of 26 | Official leaderboard | |
| Parametric CAD Bench | 50.9% | #4 of 11 | Official leaderboard | |
| TaxCalcBench v2 | 50% | #10 of 26 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 6.67% | #31 of 80 | Official leaderboard | |
| Clipping Bench | $734 | #8 of 19 | Official leaderboard | |
| OfficeQA Pro | 66.9% | #6 of 14 | Model card | |
| PolicyBench | 83.6% | #23 of 50 | Official leaderboard | |
| RealmPathology | 75.2% | #6 of 13 | Official leaderboard | |
| OR-License | 77.4% | #5 of 10 | Official leaderboard | |
| HealthBench | 57.8% | #18 of 30 | Model card | |
| HealthBench Professional | 59.8% | #16 of 25 | Model card | |
| PharmaBench (micro1) | 9.70% | #10 of 12 | Official leaderboard | |
| BenchCAD | 36.6% | #9 of 10 | Model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ZeroBench | 26% | #5 of 66 | Official leaderboard | |
| VoxelBench | 2,297 | #5 of 56 | Official leaderboard | |
| Vision Arena | 1,288 | #13 of 136 | Official leaderboard | |
| ClockBench | 65.5% | #7 of 37 | Official leaderboard | |
| Drone-Bench | 86.9% | #4 of 19 | Official leaderboard | |
| Furniture Assembly Benchmark | 60.8% | #8 of 35 | Epoch AI | |
| Code2Video Benchmark | 1,543 | #6 of 19 | Official leaderboard | |
| GDP.pdf | 24% | #15 of 39 | Epoch AI | |
| Humanity's Sixth Sense | 30.5% | #16 of 29 | Official leaderboard | |
| Blueprint-Bench 2 | 30.4% | #20 of 35 | Official leaderboard | |
| Chartography | 29.6% | #12 of 12 | Model card |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Human Pathogen Capabilities Test v2 (HPCT v2) | 80% | #1 of 71 | Official leaderboard | |
| VariantBench | 52.8% | #1 of 27 | Official leaderboard | |
| OncologyBench | 40% | #4 of 30 | Official leaderboard | |
| TxBench-Preclinical-Pharmacology | 63% | #4 of 28 | Official leaderboard | |
| BioSecBench-Surveillance | 51.5% | #4 of 27 | Official leaderboard | |
| MetagenomicsBench | 54% | #4 of 26 | Official leaderboard | |
| Anthropic undermining liberal democracy (behavioral audit metric) | 1.13 | #1 of 6 | Model card | |
| PCT (Polarized Contrastive Pairs) | 36.7% | #8 of 46 | Official leaderboard | |
| CyberGym-E2E | 89% | #1 of 5 | Official leaderboard | |
| Virology Capabilities Test (VCT) | 59% | #6 of 28 | Model card | |
| BioSecBench-Function | 44.8% | #6 of 27 | Official leaderboard | |
| scBench-Long | 27.3% | #6 of 27 | Official leaderboard | |
| Malicious computer use (Anthropic) | 93.8% | #3 of 13 | Model card | |
| DeepsecBench | 32.3 | #7 of 28 | Official leaderboard | |
| TxBench-Oligo-Discovery | 48.4% | #7 of 28 | Official leaderboard | |
| UK AISI cyber ranges | 80% | #1 of 4 | Model card | |
| FORTRESS | 16.8% | #17 of 66 | Official leaderboard | |
| SRE-Bench | 12.2% | #9 of 33 | Official leaderboard | |
| DrugDiscoveryBench | 53.7% | #6 of 19 | Official leaderboard | |
| Prompt injection live bug bounty (Anthropic with Gray Swan) | 0.08% | #1 of 3 | Model card | |
| Uncritically reporting flawed results (Anthropic) | 9.83 | #1 of 3 | Model card | |
| CyScenarioBench | 33.7% | #5 of 14 | Model card | |
| SpatialBench-Long | 30.9% | #10 of 27 | Official leaderboard | |
| SHADE-Arena | 12% | #6 of 16 | Model card | |
| MASK | 94.8% | #8 of 21 | Model card | |
| TxBench-Antibody-Discovery | 48.7% | #11 of 28 | Official leaderboard | |
| Firefox 147 exploit evaluation (Anthropic) | 52.4% | #2 of 5 | Model card | |
| EpiBench | 25.2% | #11 of 27 | Official leaderboard | |
| ExploitGym | 22% | #8 of 18 | Official leaderboard | |
| ExploitBench | 70% | #9 of 19 | Model card | |
| Malicious use of Claude Code (Anthropic) | 89% | #8 of 16 | Model card | |
| Binary Exploitation Benchmark (Anthropic, OSS-Fuzz based) | 79.4% | #3 of 6 | Model card | |
| UK AISI sabotage continuation evaluation | 0.10% | #2 of 4 | Model card | |
| Overconfidence: verifying command-line syntax (Anthropic) | 4.90 | #1 of 2 | Model card | |
| Secure and defensive coding (Anthropic) | 0.61% | #1 of 2 | Model card | |
| Single-turn benign request evaluations (Anthropic) | 0.09% | #10 of 18 | Model card | |
| BioSecBench-Refusal | 31.8% | #13 of 22 | Official leaderboard | |
| DistressBench | 70.2% | #25 of 42 | Official leaderboard | |
| DNA Synthesis Screening Evasion | 70% | #3 of 5 | Model card | |
| Prompt injection robustness against adaptive attackers, coding (Anthropic) | 0.56% | #8 of 13 | Model card | |
| Prompt injection robustness against adaptive attackers, computer use (Anthropic) | 0.54% | #8 of 13 | Model card | |
| CyberBench v1.1 (Vals) | 65.4% | #27 of 43 | Official leaderboard | |
| BBQ (Bias Benchmark for QA) | 81.6% | #16 of 24 | Model card | |
| Gray Swan IPI Arena benchmark | 0.20% | #7 of 10 | Model card | |
| Cyber harm coverage (Anthropic) | 99.6% | #3 of 4 | Model card | |
| Anthropic political even-handedness eval | 96.3% | #10 of 13 | Model card | |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 3.84% | #10 of 13 | Model card | |
| Rewind-attacker cyber adversarial robustness (Anthropic) | 5.60% | #4 of 5 | Model card | |
| UK AISI unprompted sabotage evaluation | 0% | #4 of 5 | Model card | |
| Single-turn harmful request evaluations (Anthropic) | 96.3% | #15 of 18 | Model card | |
| Anthropic automated behavioral audit | 2.30 | #5 of 6 | Model card | |
| LinuxArena | 0.70% | #5 of 6 | Model card | |
| CyberBench (Vals) | 40.7% | #24 of 26 | Official leaderboard | |
| Lazy investigation (Anthropic) | 9.94 | #1 of 1 | Model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LMCA | 64.5% | #4 of 139 | Epoch AI | |
| ACCoRD | 82.0 | #5 of 150 | Official leaderboard | |
| Conceptual Reasoning Index | 73.6 | #5 of 149 | Official leaderboard | |
| Political Science LLM Benchmark | 71.4% | #1 of 29 | Official leaderboard | |
| Vals Multimodal Index | 73.9% | #2 of 33 | Official leaderboard | |
| CAIS Text Capabilities Index | 52.2 | #7 of 66 | Official leaderboard | |
| CAIS Risk Index | 36.1 | #9 of 65 | Official leaderboard | |
| CAIS Vision Capabilities Index | 68.0 | #9 of 53 | Official leaderboard | |
| SAGE (Vals) | 49.4% | #21 of 81 | Official leaderboard | |
| Vals Science Index | 47.9% | #7 of 23 | Official leaderboard | |
| Anthropic ECI (AECI) | 162.1 | #4 of 6 | Model card |
Anthropic's models over time
Anthropic's models in color, against every other model on the Epoch Capabilities Index, by release date.
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 5 24 Jul 2026
Capabilities
- SWE-bench Verified 96
- SWE-Bench Pro 79.2
- SWE-bench Multilingual 89.5
- SWE-bench Multimodal 59.4
- DeepSWE 68.8
- FrontierCode 1.1 53.4
- Terminal-Bench 3.0 43.3
- IMO 2026 100
- Riemann-bench 60
- ArXivMath 90.8
- ProgramBench 93
- Humanity's Last Exam 56.3
- BrowseComp 90.8
- DeepSearchQA 95
- DRACO 88.6
- Chartography 29.6
- BenchCAD 36.6
- OSWorld 2.0 70.6
- GDP.pdf 83.4
- OfficeQA 78.1
- OfficeQA Pro 66.9
- MCP Atlas 85.8
- Harvey LAB (Legal Agent Benchmark) 23.6
- GDPval-AA v2 1861
- AA-Briefcase 1720
- Toolathlon-Verified 80.6
- AutomationBench 26
- ARC-AGI 97.5
- ARC-AGI-2 90.4
- ARC-AGI-3 30.2
- Global MMLU 92.5
- MILU 92.1
- INCLUDE 89.8
- BioMysteryBench
- SpatialBench (LatchBio) 72.5
- scBench (LatchBio SingleCellBench) 60.6
- ProteinGym 47.7
- Protein Design (Anthropic) 42.5
- Organic Chemistry V2 (Anthropic) 61.6
- Protocols (Anthropic life sciences)
Health
- HealthBench 57.8
- HealthBench Professional 59.8
Hallucination
Safety
- Single-turn harmful request evaluations (Anthropic) 96.3
- Single-turn benign request evaluations (Anthropic) 0.09
- 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) 89
- Malicious computer use (Anthropic) 93.8
Robustness
- Rewind-attacker cyber adversarial robustness (Anthropic) 5.6
- Gray Swan IPI Arena benchmark 0.2
- Prompt injection live bug bounty (Anthropic with Gray Swan) 0.08
- Prompt injection robustness against adaptive attackers, coding (Anthropic) 0.56
- Prompt injection robustness against adaptive attackers, computer use (Anthropic) 0.54
- Prompt injection robustness against adaptive attackers, browser use (Anthropic) 3.84
Alignment
- Anthropic automated behavioral audit 2.3
- Anthropic undermining liberal democracy (behavioral audit metric) 1.13
- UK AISI unprompted sabotage evaluation 0
- UK AISI sabotage continuation evaluation 0.1
- MASK 94.8
- Uncritically reporting flawed results (Anthropic) 9.83
- Overconfidence: verifying command-line syntax (Anthropic) 4.9
- Lazy investigation (Anthropic) 9.94
Monitorability
- SHADE-Arena 12
- LinuxArena 0.7
Bio and chem risk
Cyber risk
- ExploitBench 70
- Binary Exploitation Benchmark (Anthropic, OSS-Fuzz based) 79.4
- Firefox 147 exploit evaluation (Anthropic) 52.4
- CyScenarioBench 33.7
- ExploitGym 22
- UK AISI cyber ranges 80
- Vulnerability finding evaluations (Anthropic)
- Secure and defensive coding (Anthropic) 0.61
- Cyber harm coverage (Anthropic) 99.6