- Agentic coding FrontierCode 2.0, Xhigh effort, best effort score (launch table and Table 8.1.A) 49.1
- Agentic coding FrontierSWE v2, max effort (launch table and Table 8.1.A) 69.9
- Business workflows AutomationBench, max effort, default fallbacks (launch table and Table 8.1.A) 51.4
- Agentic scientific research Terminal-Bench-Science 0.1, Fable 5.2 with safeguards, max effort (launch table) 74.6
- Agentic search WANDR (relaxed score), max effort (launch table and Table 8.1.A) 82.1
- Computer use OSWorld 2.1 (offline set, partial), max effort (launch table and Table 8.1.A) 89.3
- Robotics RoboDojo Sim, internally run, Max effort (launch table and robotics text) 64.9
- Knowledge work AA-Briefcase v1.1, 1785 Elo at max effort. Rubric pass rate 66.1% (launch table and Table 8.1.A) 1785
- Knowledge work Remote Labor Index (RLI), High effort (launch table and Table 8.1.A) 41.3
- Multidisciplinary reasoning Humanity's Last Exam (with tools), max effort. No tools 73.2% (launch table and Table 8.1.A) 74.2
- Visual chart recognition Chartography (no tools), max effort. With tools 91.6% (launch table and Table 8.1.A) 73.4
- GDPval-AA v2.1, Elo at max effort (Table 8.1.A only) 1831
- SWE-bench Pro, adaptive thinking max effort, average of five trials (Table 8.1.A only) 93
- SWE-bench Multilingual, adaptive thinking max effort, average of five trials (Table 8.1.A only) 96.6
- SWE-bench Multimodal, adaptive thinking max effort, average of five trials (Table 8.1.A only) 74.1
Claude Fable 5.2
Claude Fable 5.2 ranks first on 51 of the 139 benchmarks with five or more models scored.
150
Benchmarks with a score
51
Ranked first, out of five or more models
96
In the top three
Anthropic Nov 2026
Claude Fable 5.2
Closed weights Flagship
173.4
#1 of 256
Epoch Capabilities Index
Self-Reported Index 175.3 #1 of 404
- AutomationBench 51.5% #2 /42
- GPQA Diamond 93.9% #18 /229
- HealthBench Professional 67.7% #3 /26
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 | |
|---|---|---|---|---|
| Humanity's Last Exam | 74.2% | #1 of 154 | Lab model card | |
| SimpleQA Verified | 94% | #1 of 94 | Epoch AI run | |
| HLE-Diamond | 77.9% | #1 of 20 | Official leaderboard | |
| Extended NYT Connections | 97.3% | #7 of 103 | Official leaderboard | |
| Global MMLU | 96% | #1 of 14 | Lab model card | |
| MILU | 94.1% | #1 of 11 | Lab model card | |
| BullshitBench V2 | 67.3% | #24 of 129 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DTBench | 99.2% | #1 of 180 | Official leaderboard | |
| SimpleBench | 94.9% | #1 of 106 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 95.1% | #1 of 86 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 100% | #2 of 68 | Official leaderboard | |
| BioMysteryBench | 83.7% | #1 of 30 | Official leaderboard | |
| ARC-AGI-2 | 93.8% | #4 of 98 | Official leaderboard | |
| MysteryMechanism | 51.8% | #2 of 31 | Official leaderboard | |
| ProofBench v1.1 | 100% | #4 of 52 | Official leaderboard | |
| GPQA Diamond | 93.9% | #18 of 229 | Epoch AI run | |
| OTIS Mock AIME 2024-2025 | 99.7% | #17 of 212 | Epoch AI run | |
| Terminal-Bench-Science | 74.6% | #1 of 12 | Lab model card | |
| ARC-AGI | 98.5% | #8 of 95 | Official leaderboard | |
| De novo protein binder design (Anthropic) | 85.1% | #1 of 3 | Lab model card | |
| Fermat's Last Theorem in Lean (Anthropic) | 35% | #1 of 1 | Lab model card | |
| STEM Professor Bench | 61.8% | #1 of 1 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Code Arena | 1,836 | #1 of 128 | Official leaderboard | |
| Vibe Code Bench | 93.8% | #1 of 120 | Official leaderboard | |
| SWE-Bench Pro | 93% | #1 of 81 | Lab model card | |
| SWE-bench Multilingual | 96.6% | #1 of 73 | Lab model card | |
| DesignArena 3D Design | 1,516 | #2 of 136 | Official leaderboard | |
| DesignArena React Native | 1,321 | #1 of 56 | Official leaderboard | |
| DesignArena Website | 1,352 | #3 of 151 | Official leaderboard | |
| ALE-Bench | 2,986 | #3 of 126 | Official leaderboard | |
| FrontierCode 2.0 | 50.2% | #1 of 32 | Official leaderboard | |
| DesignArena Code | 1,374 | #5 of 143 | Official leaderboard | |
| SWE Atlas v1.1 | 80.5% | #1 of 28 | Official leaderboard | |
| Vals RSI Index | 0.38 | #1 of 27 | Official leaderboard | |
| WeirdML v3 | 43% | #1 of 27 | Official leaderboard | |
| FrontierSWE | 69.9% | #1 of 25 | Official leaderboard | |
| DesignArena Game Dev | 1,399 | #7 of 143 | Official leaderboard | |
| DesignArena UI Component | 1,357 | #7 of 139 | Official leaderboard | |
| DesignArena Agentic Overall | 1,354 | #4 of 74 | Official leaderboard | |
| sudo L7 | 52.2% | #1 of 18 | Official leaderboard | |
| Code Migration v1.1 | 83.3% | #2 of 30 | Official leaderboard | |
| OWE-Bench | 54.6% | #1 of 14 | Official leaderboard | |
| SWE-bench Multimodal | 74.1% | #1 of 12 | Lab model card | |
| CursorBench | 60% | #2 of 21 | cursor.com, via Epoch AI | |
| CursorBench 4.0 | 60% | #2 of 21 | Official leaderboard | |
| PostTrainBench 1.2 | 44.6% | #2 of 21 | Official leaderboard | |
| Web Clone Bench | 62% | #1 of 10 | Official leaderboard | |
| DeepSWE | 74.3% | #8 of 56 | Lab model card | |
| DAXBench v2 | 87.8% | #8 of 47 | Official leaderboard | |
| DesignArena Image to HTML | 1,247 | #9 of 51 | Official leaderboard | |
| MacroscopeBench | 78.5% | #5 of 16 | Official leaderboard | |
| ProgramBench tests passed | 85.6% | #4 of 10 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Chess Puzzles | 79% | #1 of 93 | Epoch AI run | |
| Mystery Game Puzzles | 90% | #1 of 84 | Epoch AI run | |
| DuelLab GameBench 2 | 99.1 | #1 of 57 | Official leaderboard | |
| Agent Arena | 30.4% | #1 of 56 | Official leaderboard | |
| Game Arena Dots and Boxes | 1,393 | #1 of 40 | Official leaderboard | |
| Game Arena Go | 1,877 | #1 of 39 | Official leaderboard | |
| RuneBench | 6.98 | #2 of 76 | Official leaderboard | |
| EBR-bench | 90.5% | #1 of 31 | Epoch AI run | |
| Terminal-Bench 4.1 | 65.8% | #2 of 43 | Official leaderboard | |
| AutomationBench | 51.5% | #2 of 42 | Official leaderboard | |
| Kaggle Game Arena | 546 | #2 of 42 | Official leaderboard | |
| Game Arena Lines of Action | 1,117 | #2 of 39 | Official leaderboard | |
| Game Arena Nine Men's Morris | 1,177 | #2 of 39 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 654 | #2 of 39 | Official leaderboard | |
| Game Arena Four in a Row | 819 | #3 of 43 | Official leaderboard | |
| Terminal-Bench 4.0 | 67.6% | #2 of 28 | Lab model card | |
| Game Arena Checkers | 1,326 | #3 of 41 | Official leaderboard | |
| Game Arena Bargaining | 252 | #3 of 39 | Official leaderboard | |
| Game Arena Word Art | 2,151 | #4 of 39 | Official leaderboard | |
| ARC-AGI-3 | 97.7% | #3 of 27 | Official leaderboard | |
| Game Arena Word Association | 444 | #5 of 40 | Official leaderboard | |
| Game Arena Clobber | 640 | #5 of 39 | Official leaderboard | |
| Game Arena Reversi | 882 | #5 of 39 | Official leaderboard | |
| Game Arena Chess | 1,313 | #8 of 42 | Official leaderboard | |
| SpireClimb StS1 A20 Heart | 21.7% | #2 of 9 | Official leaderboard | |
| DRACO | 87.7% | #3 of 12 | Lab model card | |
| WANDR | 82.1% | #1 of 4 | Lab model card | |
| Game Arena Dark Hex | 319 | #12 of 41 | Official leaderboard | |
| OSWorld 2.1 (offline subset) | 89.3% | #1 of 3 | Lab model card | |
| Game Arena Coin Game | 395 | #15 of 39 | Official leaderboard | |
| SpireClimb | 19.6% | #4 of 9 | Epoch AI run | |
| SpireClimb StS2 A10 | 17.6% | #4 of 9 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| YC-Bench | $3,445,495 | #1 of 59 | Official leaderboard | |
| Excel Modeling Benchmark v1.1 | 77.4% | #1 of 42 | Official leaderboard | |
| Vals Finance Agent v2.1 | 69.5% | #1 of 40 | Official leaderboard | |
| Vals Index v2.2 | 71.3% | #1 of 37 | Official leaderboard | |
| APEX-Agents 1.1 | 87.6% | #2 of 65 | Official leaderboard | |
| APEX-Agents | 80.1% | #3 of 61 | Collected by Epoch AI, source not named | |
| GDPx | 16.4% | #1 of 20 | Official leaderboard | |
| Remote Labor Index | 41.3% | #1 of 19 | Official leaderboard | |
| APEX-Accounting | 70.4% | #3 of 49 | Official leaderboard | |
| TaxCalcBench v2 | 82% | #2 of 32 | Official leaderboard | |
| EEBench | 77.3% | #2 of 30 | Official leaderboard | |
| RWE-bench | 62% | #1 of 15 | Official leaderboard | |
| Immigration Bench | 74% | #2 of 25 | Official leaderboard | |
| DesignArena PDF | 1,296 | #3 of 37 | Official leaderboard | |
| HealthBench Professional | 67.7% | #3 of 26 | Lab model card | |
| OfficeQA Pro | 72.9% | #2 of 16 | Lab model card | |
| OfficeQA | 83.4% | #2 of 14 | Lab model card | |
| PharmaBench (micro1) | 23.3% | #2 of 13 | Official leaderboard | |
| Company Agent Bench | 82.2% | #1 of 6 | Official leaderboard | |
| Vals Finance Index | 63.6% | #7 of 40 | Official leaderboard | |
| BenchCAD 2.1 | 83.5% | #1 of 5 | Official leaderboard | |
| Legal Research Bench | 46.1% | #22 of 93 | Official leaderboard | |
| LegalReasonBench | 49.8% | #3 of 10 | Official leaderboard | |
| MissionReport | 63.8% | #3 of 10 | Official leaderboard | |
| PRBench | 49% | #15 of 48 | Official leaderboard | |
| Tax Agent Bench | 32.1% | #27 of 84 | Official leaderboard | |
| PhysicianBench | 64.4% | #1 of 3 | Lab model card | |
| Real-World Finance v2 (Anthropic) | 1,637 | #1 of 3 | Lab model card | |
| AA-Briefcase v1.1 | 1,785 | #3 of 8 | Lab model card | |
| GDPval-AA v2.1 | 1,831 | #4 of 10 | Lab model card | |
| HealthBench | 59.6% | #13 of 31 | Lab model card | |
| Vals Legal Index | 69.2% | #21 of 38 | Official leaderboard | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 5.00% | #31 of 38 | Official leaderboard | |
| Harvey LAB-AA v1.1 | 7.20% | #3 of 3 | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BVB | 76.4 | #1 of 45 | Official leaderboard | |
| Furniture Assembly Benchmark | 90% | #1 of 35 | Official leaderboard | |
| Blueprint-Bench 2 | 56.6% | #2 of 38 | Official leaderboard | |
| Humanity's Sixth Sense | 59.5% | #2 of 34 | Official leaderboard | |
| MineBench | 2,102 | #5 of 75 | Official leaderboard | |
| Dual-Arm Sim | 59.8 | #1 of 11 | Official leaderboard | |
| Embodied First Exam | 49.6% | #1 of 9 | Official leaderboard | |
| Chartography | 73.4% | #6 of 14 | Lab model card | |
| RoboDojo | 64.9% | #1 of 1 | Lab model card |
Safety and security
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ACCoRD | 89.4 | #1 of 152 | Official leaderboard | |
| Conceptual Reasoning Index | 79.0 | #1 of 151 | Official leaderboard | |
| LMCA | 70.6% | #1 of 147 | Official leaderboard | |
| CAIS Risk Index | 8.40 | #1 of 67 | Official leaderboard | |
| CAIS Text Capabilities Index | 66.3 | #2 of 68 | Official leaderboard | |
| LiveBench | 84.9% | #1 of 31 | Official leaderboard | |
| Vals Science Index | 63.7% | #2 of 25 | Official leaderboard |
Its lab's numbers put it at 175.3, 1.9 points above its ECI
The Self-Reported Index fits one score to the 18 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 172.3 to 178.3. How we compute it.
- SWE-bench Multilingual 96.6
- Global MMLU 96
- MILU 94.1
- SWE-Bench Pro 93
- DRACO 87.7
- ProgramBench tests passed 85.6
- OfficeQA 83.4
- Terminal-Bench-Science 74.6
- DeepSWE 74.3
- Humanity's Last Exam 74.2
- SWE-bench Multimodal 74.1
- Chartography 73.4
- OfficeQA Pro 72.9
- FrontierSWE 69.9
- HealthBench Professional 67.7
- Terminal-Bench 4.0 67.6
- HealthBench 59.6
- AutomationBench 51.4
Its lab's numbers next to 7 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 |
|---|---|---|---|
| ARC-AGI-3 | 79.3% ARC-AGI-3, verified headline Standard harness score, High effort. Provider Adapter 97.72% at High. At Max: Standard 65.66%, Provider Adapter 94.90% (Section 8.15) | 97.7% Leaderboard, high, Provider Adapter | −18 pts |
| BenchCAD 2.1 | 75.4% BenchCAD 2.1 Core, 100 problems, no tools, adaptive thinking max effort, official scorer, five runs, up to 60 rounds and 500k output tokens. With tools 83.5%, without round or token limits. Not comparable with BenchCAD 1.0 (Section 8.13.2) | 83.5% Leaderboard, max, Claude Code | −8.1 pts |
| FrontierCode 2.0 | 49.1% Agentic coding FrontierCode 2.0, Xhigh effort, best effort score (launch table and Table 8.1.A) | 50.2% Leaderboard, xhigh, Claude Code | −1.1 pts |
| AutomationBench | 51.4% Business workflows AutomationBench, max effort, default fallbacks (launch table and Table 8.1.A) | 51.5% Leaderboard, max, default fallbacks | Same |
| Remote Labor Index | 41.3% Knowledge work Remote Labor Index (RLI), High effort (launch table and Table 8.1.A) | 41.3% Leaderboard | Same |
| FrontierSWE | 69.9% Agentic coding FrontierSWE v2, max effort (launch table and Table 8.1.A) | 69.9% Leaderboard, mean@5, proximus | Same |
| CursorBench 4.0 | 60% CursorBench 4.0, max effort, independently run by Cursor. Additional prompt and harness improvements may prevent comparisons with earlier system cards (Section 8.8) | 60% Leaderboard, max, Cursor | 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. 41 of them aren't in the launch post.
Capabilities
- SWE-Bench Pro 93
- SWE-bench Multilingual 96.6
- SWE-bench Multimodal 74.1
- DeepSWE 74.3
- FrontierCode 2.0 49.1
- Terminal-Bench 4.0 67.6
- Terminal-Bench-Science 74.6
- FrontierSWE 69.9
- CursorBench 4.0 60
- STEM Professor Bench 61.8
- ProgramBench tests passed 85.6
- Humanity's Last Exam 73.2
- DRACO 87.7
- WANDR 82.1
- Fermat's Last Theorem in Lean (Anthropic) 35
- Chartography 73.4
- BenchCAD 2.1 75.4
- OSWorld 2.1 (offline subset) 89.3
- OfficeQA 83.4
- OfficeQA Pro 72.9
- Harvey LAB-AA v1.1 7.2
- GDPval-AA v2.1 1831
- AA-Briefcase v1.1 1785
- AutomationBench 51.4
- Real-World Finance v2 (Anthropic) 1637
- Remote Labor Index 41.3
- ARC-AGI-3 79.3
- Global MMLU 96
- MILU 94.1
- De novo protein binder design (Anthropic) 85.1
Health
- HealthBench 59.6
- HealthBench Professional 67.7
- PhysicianBench 64.4
Safety
- Single-turn harmful request evaluations (Anthropic) 93.7
- Single-turn benign request evaluations (Anthropic) 0
- 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) 82.8
- Malicious computer use (Anthropic) 88.4