- Agents' Last Exam v1.1 (GPT-6.2 Astra Xhigh, 77.6 at Max) 78.3
- Terminal-Bench-Science 0.1 (GPT-6.2 Astra Max) 77.1
- Artificial Analysis Intelligence Index v5.0.1 (GPT-6.2 Astra Max), cited from Artificial Analysis 62.1
- FrontierCode 2.0 (GPT-6.2 Astra Max) 44.2
- LifeVal (GPT-6.2 Astra Max) 86.1
- Harvey-LAB-AA v1.1 (GPT-6.2 Astra Max) 15
- GDPval v2 (Internal) (GPT-6.2 Astra Max) 50.4
- AA-AnalystAgent, questions answered correctly on all five independent attempts 62.5
- AA-3D, 1,318 Elo 1318
- Vals Science Index v1 (GPT-6.2 Astra Max) 68.4
- BenchCAD 2.1 Core (GPT-6.2 Astra Max) 82.5
- AutomationBench 1.0.6 (GPT-6.2 Astra Max) 52.4
- ExploitBench Internal Port (June to August 2026), success rate (GPT-6.2 Astra Max, safeguards and classifiers disabled) 53.5
- SecureSWE (GPT-6.2 Astra Max) 65.7
- PatchBench (GPT-6.2 Astra Xhigh and Max) 79.7
GPT-6.2 Astra
GPT-6.2 Astra ranks first on 34 of the 157 benchmarks with five or more models scored.
GPT-6.2 Astra
Closed weights Flagship
- AutomationBench 52.4% #3 /45
- MultiChallenge 73.2% #9 /45
- FrontierCode 2.0 44.2% #4 /24
Where has it been number one?
GPT-6.2 Astra is number one on 21 benchmarks today. It also led 1 more before a newer model passed it.
Sources: Official leaderboards and Epoch AI runs (CC BY 4.0)
How does it do against its rivals?
GPT-6.2 Astra beats Claude Fable 5 on 31 of the 45 benchmarks they share. Claude Fable 5.2 does best against it, ahead on 56 of 101.
- Claude Fable 5 31 ahead, 12 behind, 2 tied Compare
-
Gemini 4 Argon 65 ahead, 33 behind, 5 tied Compare
- Claude Opus 5.5 69 ahead, 36 behind, 7 tied Compare
-
Gemini 4 Hydrogen 51 ahead, 27 behind, 2 tied Compare
-
Gemini 4 Carbon 46 ahead, 45 behind, 4 tied Compare
- Claude Fable 5.2 40 ahead, 56 behind, 5 tied Compare
Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
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 | |
|---|---|---|---|---|
| Angry Tweets | 67% | #1 of 149 | Official leaderboard | |
| EuroEval | 1.17 | #1 of 75 | Official leaderboard | |
| Lech Mazur Creative Writing | 4.30 | #1 of 55 | Official leaderboard | |
| SimpleQA Verified | 78.5% | #2 of 96 | Epoch AI run | |
| Fiction Bench (Story Hacker) | 73.6 | #1 of 36 | Official leaderboard | |
| FoSent | 78% | #5 of 147 | Official leaderboard | |
| Extended NYT Connections | 98.2% | #5 of 109 | Official leaderboard | |
| The Aggregate Chart Recall | 0.22 | #1 of 18 | Official leaderboard | |
| HLE-Diamond | 72.3% | #2 of 22 | Official leaderboard | |
| EuroEval ERRNews | 31.7 | #7 of 76 | Official leaderboard | |
| LMArena (Elo) | 1,468 | #47 of 345 | Official leaderboard | |
| The Aggregate Day-one Recall Check | 48.5% | #5 of 36 | Official leaderboard | |
| LLM Debate Benchmark | 1,673 | #9 of 58 | Official leaderboard | |
| The Aggregate Personal Style Imitation | 75.5% | #3 of 17 | Official leaderboard | |
| The Aggregate Trip Outcome Prediction | 69.8% | #5 of 26 | Official leaderboard | |
| MultiChallenge | 73.2% | #9 of 45 | Official leaderboard | |
| Caliper Bench v3 | 82.0 | #17 of 71 | Official leaderboard | |
| The Aggregate Long-tail Book Recall | 56.1% | #6 of 21 | Official leaderboard | |
| The Aggregate Exotic Fact Recall | 60% | #5 of 15 | Official leaderboard | |
| BullshitBench V2 | 55% | #50 of 137 | Official leaderboard | |
| The Aggregate Asteroid Numbers | 47.9% | #6 of 10 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 96.8% | #1 of 231 | Epoch AI run | |
| ARC-AGI-2 | 95.8% | #1 of 101 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 100% | #3 of 75 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 92.6% | #4 of 93 | Official leaderboard | |
| MysteryMechanism | 53.1% | #2 of 34 | Official leaderboard | |
| FrontierMath Erdős | 10.3% | #1 of 16 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 100% | #17 of 214 | Epoch AI run | |
| BioMysteryBench | 80.7% | #3 of 33 | Official leaderboard | |
| ARC-AGI | 98.5% | #9 of 98 | Official leaderboard | |
| ArXivLean | 79.3% | #1 of 10 | Official leaderboard | |
| Autoform-Hard | 87.5% | #1 of 8 | Official leaderboard | |
| SimpleBench | 77.3% | #16 of 110 | Official leaderboard | |
| RandomBench | 0.30 | #6 of 37 | Official leaderboard | |
| Terminal-Bench-Science | 77.1% | Not ranked | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DesignArena Game Dev | 1,450 | #2 of 154 | Official leaderboard | |
| Ramp SWE-Bench v2 | 76.4% | #1 of 22 | Official leaderboard | |
| DAXBench v2 | 93.6% | #3 of 49 | Official leaderboard | |
| OWE-Bench | 67.2% | #1 of 16 | Official leaderboard | |
| Vibe Code Bench | 89.9% | #9 of 124 | Official leaderboard | |
| Vibe Code Bench 1-100 v1.1 | 55.2% | #1 of 11 | Official leaderboard | |
| FrontierCode 2.0 | 44.2% | #4 of 24 | Official leaderboard | |
| MirrorCode | 67.8% | #8 of 41 | Epoch AI run | |
| Vals RSI Index | 0.29 | #6 of 28 | Official leaderboard | |
| MacroscopeBench | 81.4% | #4 of 18 | Official leaderboard | |
| LLM 3D Worlds Arena | 1,123 | #5 of 12 | Official leaderboard | |
| Ramen Bench | 4 | #12 of 24 | Official leaderboard | |
| SWE Atlas v1.1 | 60.5% | #17 of 31 | Official leaderboard | |
| MLE-bench | 96.1% | #1 of 12 lab numbers | Lab model card | |
| Internal Research Debugging Evaluation | 81.8% | #1 of 8 lab numbers | Lab model card | |
| KernelGen 1P | 76.9% | #1 of 6 lab numbers | Lab model card | |
| AA-3D | 1,318 | #1 of 3 lab numbers | Lab model card | |
| NanoGPT Speedrun | 42.9% | #1 of 2 lab numbers | Lab model card | |
| OpenAI Epoch Capabilities Index (OECI) | 170.1 | #1 of 2 lab numbers | Lab model card | |
| PostTrainBench | 74.8% | Not ranked | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Mystery Game Puzzles | 95% | #1 of 86 | Epoch AI run | |
| RuneBench | 7.85 | #1 of 79 | Official leaderboard | |
| SnakeBench | 38 | #5 of 201 | Official leaderboard | |
| EBR-bench | 92.4% | #1 of 34 | Epoch AI run | |
| Chess Puzzles | 78% | #3 of 95 | Epoch AI run | |
| Dubesor Lichess Puzzle Rating | 2,244 | #2 of 62 | Official leaderboard | |
| ARC-AGI-3 | 100% | #1 of 31 | Official leaderboard | |
| KradleBench | 78.3 | #1 of 26 | Official leaderboard | |
| Dubesor AI Chess Leaderboard | 1,731 | #12 of 250 | Official leaderboard | |
| Crest Arcade Chess | 1,758 | #1 of 19 | Official leaderboard | |
| Crest Arcade Hustle Sim | 83.7 | #1 of 17 | Official leaderboard | |
| AutomationBench | 52.4% | #3 of 45 | Official leaderboard | |
| SpireClimb | 41.3% | #1 of 12 | Epoch AI run | |
| SpireClimb StS2 A10 | 49.3% | #1 of 12 | Official leaderboard | |
| RedstoneBench | 88.7 | #1 of 10 | Official leaderboard | |
| Maze Bench | 17% | #6 of 58 | Official leaderboard | |
| Portal2Bench | 26.0 | #1 of 9 | Official leaderboard | |
| Crest Arcade Sudoku | 100% | #2 of 17 | Official leaderboard | |
| Terminal-Bench 4.1 | 62.1% | #6 of 42 | Official leaderboard | |
| Crest Arcade Nonograms | 69.4% | #2 of 14 | Official leaderboard | |
| The Aggregate Inversion and Execution Ladder | 100% | #3 of 20 | Official leaderboard | |
| Toloka Arena Enterprise Tool Use (November 2026) | 72% | #7 of 43 | Official leaderboard | |
| SpireClimb StS1 A20 Heart | 33.3% | #2 of 12 | Official leaderboard | |
| Crest Arcade Elemental Armory | 58.7% | #3 of 17 | Official leaderboard | |
| Crest Arcade Mines | 83.3% | #3 of 17 | Official leaderboard | |
| The Aggregate Puzzle Portfolio Ladder | 96.7% | #5 of 24 | Official leaderboard | |
| Crest Arcade Ravelin | 89.2 | #3 of 14 | Official leaderboard | |
| Crest Arcade Wildhaven | 85.6% | #3 of 14 | Official leaderboard | |
| Crest Arcade Micro Mouse | 79.0 | #4 of 17 | Official leaderboard | |
| Crest Arcade FreeCell | 33.3% | #5 of 17 | Official leaderboard | |
| The Aggregate Lichess Rating Prediction | 354 | #10 of 19 | Official leaderboard | |
| PokéBench: Tool Assisted Benchmark | 10.0 | #19 of 34 | Official leaderboard | |
| Crest Arcade Nations Royale | 83.3 | #11 of 16 | Official leaderboard | |
| Crest Arcade Starjar | 98.7 | #10 of 14 | Official leaderboard | |
| The Aggregate Lichess Pairwise Strength | 76% | #15 of 19 | Official leaderboard | |
| LifeVal | 86.1% | #1 of 2 lab numbers | Lab model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Construction's Last Exam | 58.1% | #1 of 30 | Official leaderboard | |
| DAYJOB: Law | 41.4% | #1 of 22 | Official leaderboard | |
| Tuesday Work Index | 70.3 | #3 of 56 | Official leaderboard | |
| Agents' Last Exam v1.1 | 47.5% | #1 of 18 | Official leaderboard | |
| ContractorOS Quantity Takeoff | 53.8% | #3 of 43 | Official leaderboard | |
| DeepFinance | 61.4% | #2 of 28 | Official leaderboard | |
| DesignArena PDF | 1,282 | #4 of 48 | Official leaderboard | |
| YC-Bench | $2,601,027 | #6 of 66 | Official leaderboard | |
| APEX-Accounting | 69.8% | #5 of 52 | Official leaderboard | |
| Wojak-Bench 3D | 1,273 | #1 of 10 | Official leaderboard | |
| APEX-Agents 1.1 | 83.9% | #7 of 69 | Official leaderboard | |
| VetEval | 71.4 | #3 of 29 | Official leaderboard | |
| Excel Modeling Benchmark v1.1 | 76% | #5 of 46 | Official leaderboard | |
| DAYJOB: Accounting | 42% | #3 of 26 | Official leaderboard | |
| EEBench | 74.6% | #4 of 34 | Official leaderboard | |
| Argo-Bench | 62.9 | #3 of 24 | Official leaderboard | |
| APEX-1 | 71.1% | #11 of 75 | Official leaderboard | |
| TaxCalcBench v2 | 76% | #6 of 38 | Official leaderboard | |
| Deta Studio AI Construction Takeoff | 73 | #3 of 17 | Official leaderboard | |
| Vals Index v2.2 | 67.7% | #8 of 41 | Official leaderboard | |
| GDP.xlsx | 29.7% | #7 of 34 | Official leaderboard | |
| HW-Bench | 53% | #4 of 19 | Official leaderboard | |
| GradeBench | 56.3% | #4 of 18 | Official leaderboard | |
| BenchCAD 2.1 | 82.5% | #2 of 9 | Official leaderboard | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 16.3% | #12 of 42 | Official leaderboard | |
| DataAnalyticsBench | 41.0 | #15 of 51 | Official leaderboard | |
| Legal Research Bench | 42.8% | #31 of 97 | Official leaderboard | |
| Ramp Accounting Bench | 50.4% | #6 of 18 | Official leaderboard | |
| Company Agent Bench | 79.7% | #3 of 8 | Official leaderboard | |
| Periapical Lesions Bench | 12% | #3 of 8 | Official leaderboard | |
| SimReadyBench NV100 | 99.6% | #3 of 7 | Official leaderboard | |
| Vals Finance Index | 57.3% | #23 of 44 | Official leaderboard | |
| Vals Legal Index | 69.7% | #23 of 42 | Official leaderboard | |
| Vals Finance Agent v2.1 | 60.8% | #26 of 44 | Official leaderboard | |
| Tax Agent Bench | 21.8% | #53 of 87 | Official leaderboard | |
| Epoch Automation Reports | 64.9% | #4 of 6 | Official leaderboard | |
| RCM-Bench-1 | 78.4% | #22 of 27 | Official leaderboard | |
| DualEntry Accounting AI Benchmark | 78.2% | #26 of 29 | Official leaderboard | |
| Harvey LAB-AA v1.1 | 15% | #1 of 5 lab numbers | Lab model card | |
| AA-AnalystAgent | 62.5% | #1 of 2 lab numbers | Lab model card | |
| GDPval v2 (OpenAI) | 50.4% | #1 of 2 lab numbers | Lab model card | |
| HealthBench Professional | 66.3% | Not ranked | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Chartography | 90.7% | #1 of 56 | Official leaderboard | |
| BuseyBench SVG Test | 7.40 | #5 of 261 | Official leaderboard | |
| Furniture Assembly Benchmark | 89.2% | #2 of 39 | Official leaderboard | |
| RoboDojo-RC Tier 1 | 65.7% | #1 of 12 | Official leaderboard | |
| Robo-bench | 32.7% | #1 of 11 | Official leaderboard | |
| Humanity's Sixth Sense | 63.2% | #4 of 38 | Official leaderboard | |
| HearingBench | 43% | #5 of 16 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| TxBench-Preclinical-Pharmacology | 68% | #1 of 32 | Official leaderboard | |
| DeepsecBench | 61.2 | #1 of 31 | Official leaderboard | |
| DrugDiscoveryBench | 65.5% | #4 of 27 | Official leaderboard | |
| CWE-bench 2 | 66.8% | #6 of 20 | Official leaderboard | |
| Multimodal Troubleshooting Virology | 61.8% | #2 of 29 lab numbers | Lab model card | |
| AAV Capsid Packaging Prediction (OpenAI) | 0.59 | #1 of 9 lab numbers | Lab model card | |
| SHP2 Protein Function Prediction | 0.46 | #1 of 8 lab numbers | Lab model card | |
| ExploitBench - Internal Port | 53.5% | #1 of 7 lab numbers | Lab model card | |
| CyScenarioBench | 63% | #4 of 18 lab numbers | Lab model card | |
| Respecting Warnings | 17.4% | #2 of 9 lab numbers | Lab model card | |
| OpenAI Deployment Simulation | 0.45% | #2 of 7 lab numbers | Lab model card | |
| Sandbox Bench | 36.4% | #2 of 6 lab numbers | Lab model card | |
| Phage-plasmid Co-evolution | 11.6 | #3 of 8 lab numbers | Lab model card | |
| CoT-Control | 38.6% | #5 of 11 lab numbers | Lab model card | |
| Coronavirus-ACE2 Cell-Entry Screen | 0.44 | #4 of 8 lab numbers | Lab model card | |
| Unintended Engagement with External Agent Messages | 11.3% | #4 of 8 lab numbers | Lab model card | |
| FrontierCyber | 31.3% | #2 of 4 lab numbers | Lab model card | |
| Cyber Utility | 78.1% | #1 of 2 lab numbers | Lab model card | |
| Model Spec Conversation Replay | 80.6 | #1 of 2 lab numbers | Lab model card | |
| Model Spec Targeted Cases | 87.8% | #1 of 2 lab numbers | Lab model card | |
| Multiturn Jailbreak Evaluations | 80.9% | #1 of 2 lab numbers | Lab model card | |
| PatchBench (OpenAI) | 79.7% | #1 of 2 lab numbers | Lab model card | |
| SecureSWE (OpenAI) | 65.7% | #1 of 2 lab numbers | Lab model card | |
| Declining to Exploit a Honeypot During Difficult ExploitGym Problems | 0% | #6 of 9 lab numbers | Lab model card | |
| Agent Replay Auto-Review | 12.3% | #2 of 3 lab numbers | Lab model card | |
| Math Side-tasking Monitor Evasion | 0% | #3 of 4 lab numbers | Lab model card | |
| Blocker Deception | 22.3% | #6 of 7 lab numbers | Lab model card | |
| Respecting Auto-Review | 0% | #8 of 9 lab numbers | Lab model card | |
| Broken Search Tool | 23% | #11 of 11 lab numbers | Lab model card | |
| Advanced Screening Evasion | 0.70 | #2 of 2 lab numbers | Lab model card | |
| Expected Cyber Refusals | 98.5% | #2 of 2 lab numbers | Lab model card | |
| Prompt Injection in Benchmark Environments | 99.2% | #2 of 2 lab numbers | Lab model card | |
| Prompt Injection in Model-Generated Environments | 84.7% | #2 of 2 lab numbers | Lab model card | |
| Epoch AI no-CoT serial reasoning steps | 12.3 | #1 of 1 lab number | Lab model card | |
| EUREKA-Bench (OpenAI) | 100.3 | #1 of 1 lab number | Lab model card | |
| Sequitur Research Lab Automation | 74.5 | #1 of 1 lab number | Lab model card | |
| UK AISI correlated monitor failure | 19% | #1 of 1 lab number | Lab model card | |
| UK AISI no-CoT maths competition problems | 79.2% | #1 of 1 lab number | Lab model card | |
| UK AISI simulated sandbox escape monitor evasion | 0.50% | #1 of 1 lab number | Lab model card | |
| UK AISI vulnerability discovery challenges | 88.6% | #1 of 1 lab number | Lab model card | |
| BioTIER Refuse | 79.8% | Not ranked | Lab model card | |
| ExploitBench | 100% | Not ranked | Lab model card | |
| MentalHealthBench | 53.3% | Not ranked | Lab model card | |
| nocot-bench (NCRI) | 171.0 | Not ranked | Lab model card | |
| SEC-bench Pro | 86.8% | Not ranked | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v5 | 62.1 | #2 of 69 | Official leaderboard | |
| Vals Science Index | 68.4% | #1 of 28 | Official leaderboard | |
| LiveBench | 84.9% | #3 of 36 | Official leaderboard | |
| Parallel Search Capability Leaderboard | 84.5 | #5 of 51 | Official leaderboard | |
| CAIS Text Capabilities Index | 54.7 | #8 of 70 | Official leaderboard | |
| CAIS Risk Index | 40.5 | #18 of 69 | Official leaderboard | |
| Epoch Capabilities Index | 174.0 | #1 of 1 lab number | Lab model card |
What does it cost to run?
On the 78 leaderboards that publish a price, its best run usually costs 1.8 times the median model's. On 27 of them, no model comes close to its score for less.
| Benchmark | Score | Price | From cheapest to priciest | Against the median | Close to its score for less Close for less |
|---|---|---|---|---|---|
| BuseyBench SVG Test | 7.40 | $1.42 114× the median | 114× | GPT-6.1 Terra 7.30 for $0.19 | |
| Dubesor AI Chess Leaderboard | 1,731 | $0.86 4.1× the median | 4.1× | | |
| SnakeBench | 38 | $1.33 25× the median | 25× | GPT-6 Astra 37 for $0.9 | |
| BullshitBench V2 | 55% | $0.086 18× the median | 18× | GPT-6 Luna 59% for $0.00011 | |
| Vibe Code Bench | 89.9% | $7.92 2.0× the median | 2.0× | Claude Haiku 5.5 90.4% for $6.07 | |
| ARC-AGI-2 | 95.8% | $0.5 1.1× the median | 1.1× | GPT-6.1 Sol 94.2% for $0.25 | |
| ARC-AGI | 98.5% | $0.17 0.8× the median | 0.8× | GPT-6.2 Sol 97.5% for $0.059 | |
| Legal Research Bench | 42.8% | $12.84 12× the median | 12× | GLM Legal 41.8% for $0.05 | |
| Tax Agent Bench | 21.8% | $7.90 11× the median | 11× | | |
| RuneBench | 7.85 | $10.96 5.9× the median | 5.9× | GPT-6.2 Sol 7.76 for $3.41 | |
| DesignArena Game Dev | 1,450 | $0.99 13× the median | 13× | | |
| SimpleBench | 77.3% | $0.96 per run 0.3× the median | 0.3× | GPT-6.1 Sol 82.9% for $0.39 | |
| Dubesor Lichess Puzzle Rating | 2,244 | $0.052 4.2× the median | 4.2× | None | |
| Maze Bench | 17% | $313 per run 6.1× the median | 6.1× | GPT-6.1 Sol 15% for $63.24 | |
| DataAnalyticsBench | 41.0 | $4.11 per run 1.3× the median | 1.3× | | |
| Parallel Search Capability Leaderboard | 84.5 | $199 1.6× the median | 1.6× | | |
| DesignArena PDF | 1,282 | $2.29 2.0× the median | 2.0× | | |
| AutomationBench | 52.4% | $0.83 1.5× the median | 1.5× | | |
| Excel Modeling Benchmark v1.1 | 76% | $2.80 1.8× the median | 1.8× | GPT-6.2 Sol 74.3% for $0.84 | |
| Toloka Arena Enterprise Tool Use (November 2026) | 72% | $8.22 3.7× the median | 3.7× | | |
| Vals Finance Agent v2.1 | 60.8% | $7.98 4.9× the median | 4.9× | | |
| Vals Finance Index | 57.3% | $6.23 4.5× the median | 4.5× | | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 16.3% | $7.33 2.2× the median | 2.2× | | |
| Terminal-Bench 4.1 | 62.1% | $4.27 1.0× the median | 1.0× | GPT-6.2 Sol 61.5% for $1.06 | |
| Vals Legal Index | 69.7% | $9.35 4.0× the median | 4.0× | MiMo-V2.6 Flash 69% for $0.07 | |
| Vals Index v2.2 | 67.7% | $10.68 2.9× the median | 2.9× | GPT-6.2 Sol 66.1% for $3.18 | |
| Fiction Bench (Story Hacker) | 73.6 | $8.78 per run 3.1× the median | 3.1× | GPT-6.2 Sol 72.6 for $2.37 | |
| LiveBench | 84.9% | $1.65 1.8× the median | 1.8× | None | |
| EEBench | 74.6% | $2.31 1.5× the median | 1.5× | None | |
| MysteryMechanism | 53.1% | $1.72 1.3× the median | 1.3× | Claude Fable 5.2 51.8% for $1.55 | |
| BioMysteryBench | 80.7% | $1.49 1.5× the median | 1.5× | GPT-6.1 Sol 79.6% for $0.37 | |
| ARC-AGI-3 | 100% | $4,112 per run 0.9× the median | 0.9× | GPT-6.2 Sol 98.2% for $1,494 | |
| DeepsecBench | 61.2 | $59.85 per run 4.7× the median | 4.7× | None | |
| TaxCalcBench v2 | 76% | $1.26 2.2× the median | 2.2× | GPT-6.1 Sol 74% for $0.3 | |
| Construction's Last Exam | 58.1% | $11.20 per run 6.4× the median | 6.4× | None | |
| DeepFinance | 61.4% | $4.69 1.7× the median | 1.7× | None | |
| VetEval | 71.4 | $0.0074 1.0× the median | 1.0× | None | |
| Vals RSI Index | 0.29 | $342 1.1× the median | 1.1× | | |
| Vals Science Index | 68.4% | $4.76 1.4× the median | 1.4× | None | |
| DAYJOB: Accounting | 42% | $5.31 1.5× the median | 1.5× | | |
| KradleBench | 78.3 | $1.26 3.8× the median | 3.8× | None | |
| DualEntry Accounting AI Benchmark | 78.2% | $0.38 4.2× the median | 4.2× | | |
| Argo-Bench | 62.9 | $2.17 0.9× the median | 0.9× | None | |
| FrontierCode 2.0 | 44.2% | $5.97 1.2× the median | 1.2× | GPT-6.2 Sol 43.8% for $1.29 | |
| Ramen Bench | 4 | $0.94 1.0× the median | 1.0× | GPT-5.6 Sol 4 for $0.056 | |
| RCM-Bench-1 | 78.4% | $0.44 9.5× the median | 9.5× | GPT-6 Luna 85.8% for $0.0019 | |
| DAYJOB: Law | 41.4% | $5.54 1.9× the median | 1.9× | None | |
| Ramp SWE-Bench v2 | 76.4% | $3.31 3.2× the median | 3.2× | None | |
| CWE-bench 2 | 66.8% | $4.32 2.2× the median | 2.2× | GPT-6.2 Sol 66% for $0.81 | |
| HW-Bench | 53% | $23.43 1.3× the median | 1.3× | GPT-6.1 Sol 57% for $8.42 | |
| The Aggregate Lichess Rating Prediction | 354 | $0.58 per run 1.5× the median | 1.5× | | |
| Agents' Last Exam v1.1 | 47.5% | $3.05 0.8× the median | 0.8× | None | |
| GradeBench | 56.3% | $0.041 2.1× the median | 2.1× | | |
| MacroscopeBench | 81.4% | $0.79 0.2× the median | 0.2× | GPT-6.1 Terra 81.2% for $0.35 | |
| Ramp Accounting Bench | 50.4% | $1.31 0.8× the median | 0.8× | None | |
| Deta Studio AI Construction Takeoff | 73 | $0.63 0.5× the median | 0.5× | GPT-6.1 Terra 73 for $0.28 | |
| Crest Arcade Chess | 1,758 | $1.37 1.5× the median | 1.5× | None | |
| Crest Arcade Elemental Armory | 58.7% | $0.12 2.6× the median | 2.6× | GPT-6.1 Terra 67.3% for $0.013 | |
| Crest Arcade FreeCell | 33.3% | $0.62 1.7× the median | 1.7× | Claude Haiku 5.5 63.3% for $0.071 | |
| Crest Arcade Hustle Sim | 83.7 | $1.25 3.6× the median | 3.6× | None | |
| Crest Arcade Micro Mouse | 79.0 | $0.62 1.1× the median | 1.1× | None | |
| Crest Arcade Mines | 83.3% | $1.17 2.5× the median | 2.5× | | |
| Crest Arcade Sudoku | 100% | $0.2 1.2× the median | 1.2× | None | |
| RoboDojo-RC Tier 1 | 65.7% | $10.51 4.9× the median | 4.9× | None | |
| Crest Arcade Nations Royale | 83.3 | $1.06 1.2× the median | 1.2× | GPT-6.1 Terra 91.7 for $0.29 | |
| LLM Debate Benchmark | 1,673 | $0.32 2.2× the median | 2.2× | | |
| Robo-bench | 32.7% | $19.70 per run 1.7× the median | 1.7× | None | |
| Vibe Code Bench 1-100 v1.1 | 55.2% | $45.66 3.4× the median | 3.4× | None | |
| ArXivLean | 79.3% | $5.10 0.7× the median | 0.7× | None | |
| Crest Arcade Nonograms | 69.4% | $0.25 1.4× the median | 1.4× | None | |
| Crest Arcade Ravelin | 89.2 | $0.32 2.3× the median | 2.3× | None | |
| Crest Arcade Starjar | 98.7 | $1.60 1.7× the median | 1.7× | Claude Haiku 5.5 98.7 for $0.088 | |
| Crest Arcade Wildhaven | 85.6% | $0.1 2.5× the median | 2.5× | GPT-6.1 Sol 83.6% for $0.037 | |
| BenchCAD 2.1 | 82.5% | $5.40 0.6× the median | 0.6× | None | |
| Extended NYT Connections | 98.2% | $0.015 0.5× the median | 0.5× | None | |
| Company Agent Bench | 79.7% | $0.17 1.1× the median | 1.1× | GPT-6.1 Sol 78.8% for $0.062 | |
| Periapical Lesions Bench | 12% | $0.48 1.1× the median | 1.1× | None | |
| SimReadyBench NV100 | 99.6% | $0.58 1.9× the median | 1.9× | GPT-6.1 Sol 97.7% for $0.21 |
Inside agents
Boards that list GPT-6.2 Astra inside an agent or product, such as Codex CLI 0.169.0 + GPT-6.2 Astra. Those scores belong to the agent, so we don't count them as the model's own. Its own score on the same board is beside each one.
| Benchmark | Entry | Agent's score | Its own score |
|---|---|---|---|
| InferenceBench | Codex CLI 0.169.0 + GPT-6.2 Astra | 10.8× | |
| Large Language Model Football League | Tush Push Notifications + GPT-6.2 Astra | 17.0 |
Its lab's numbers put it at 179.8, 5.7 points above its ECI
The Self-Reported Index fits one score to the 6 capability results OpenAI 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 174.3 to 185.4. How we compute it.
- MLE-bench 96.1
- Internal Research Debugging Evaluation 81.8
- Terminal-Bench-Science 77.1
- PostTrainBench 74.8
- HealthBench Professional 66.3
- AutomationBench 52.4
Who measured it?
Independent runs cover 137 of its 192 benchmarks. For the other 55, we have only OpenAI's own number or a claim Epoch AI collected.
- An independent run and a claim 7
- An independent run only 130
- A claim only 55
Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Its lab's numbers next to 8 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 |
|---|---|---|---|
| GPQA Diamond | 97% GPQA Diamond (no tools), accuracy (Table 2, 2.1.1.1.1) | 96.8% Epoch AI run, max | +0.2 pts |
| AutomationBench | 52.4% AutomationBench 1.0.6 (GPT-6.2 Astra Max) | 52.4% Leaderboard, max | Same |
| Vals Science Index | 68.4% Vals Science Index v1 (GPT-6.2 Astra Max) | 68.4% Leaderboard | Same |
| Agents' Last Exam v1.1 Score | 78.3% Agents' Last Exam v1.1 (GPT-6.2 Astra Xhigh, 77.6 at Max) | 78.3% Leaderboard, xhigh, Codex | Same |
| FrontierCode 2.0 | 44.2% FrontierCode 2.0 (GPT-6.2 Astra Max) | 44.2% Leaderboard, max, Codex | Same |
| FrontierMath Erdős | 10.3% Epoch AI FrontierMath Erdős, Lean solutions to 7 of 68 problems with a $300 budget per question (2.1.3.2.2) | 10.3% Leaderboard, max, Epoch | Same |
| Artificial Analysis Intelligence Index v5 | 62.1 Artificial Analysis Intelligence Index v5.0.1 (GPT-6.2 Astra Max), cited from Artificial Analysis | 62.1 Leaderboard, max | Same |
| BenchCAD 2.1 | 82.5% BenchCAD 2.1 Core (GPT-6.2 Astra Max) | 82.5% Leaderboard, max, Codex | Same |
OpenAI's models over time
When GPT-6.2 Astra came out, no model released before it scored higher on Epoch AI's Capabilities Index. None has since.
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. 70 of them aren't in the launch post.
Hallucination
Safety
- Biology Model Refusal Evaluation
- Biological safety evaluations (OpenAI)
- OpenAI cyber safety evaluations (incl. Daybreak Blue)
- BioTIER Refuse 79.8
- OpenAI Production Benchmarks and safe-completion evals
- U18 safety evaluations
- Agentic Safe Completions
- Image input safety evaluations
- Expected Cyber Refusals 98.5
- Cyber Utility 78.1
Robustness
Alignment
- Respecting Auto-Review 0
- Respecting Warnings 17.4
- Declining to Exploit a Honeypot During Difficult ExploitGym Problems 0
- Agent Replay Auto-Review 12.3
- Blocker Deception 22.3
- Broken Search Tool 23
- Misaligned Behavior in Realistic Work Environments
- OpenAI Deployment Simulation 0.45
- Unintended Engagement with External Agent Messages 11.3
- Verbalized metagaming
- Oversight gaming
- UK AISI Out of Scope Supply Chain Attack
- Model Spec Targeted Cases 87.8
- Model Spec Conversation Replay 80.6
- UK AISI correlated monitor failure 19
Monitorability
Bio and chem risk
- GPQA Diamond 97
- Multimodal Troubleshooting Virology 61.8
- AAV Capsid Packaging Prediction (OpenAI) 0.593
- SHP2 Protein Function Prediction 0.46
- Coronavirus-ACE2 Cell-Entry Screen 0.435
- Phage-plasmid Co-evolution 11.6
- ReproBAIT
- Advanced Screening Evasion 0.698
- Bio-Probe 0.232
- EUREKA-Bench (OpenAI) 100.3
- Sequitur Research Lab Automation 74.5
Cyber risk
- SEC-bench Pro 86.8
- ExploitBench 100
- ExploitBench - Internal Port 53.5
- Sandbox Bench 36.4
- CyScenarioBench 63
- FrontierCyber 31.3
- UK AISI vulnerability discovery challenges 88.6