- Agents' Last Exam v1.1 (GPT-6.2 Sol Max) 73
- Terminal-Bench-Science 0.1 (GPT-6.2 Sol Max) 68
- Artificial Analysis Intelligence Index v5.0.1 (GPT-6.2 Sol Max), cited from Artificial Analysis 57.3
- FrontierCode 2.0 (GPT-6.2 Sol Max) 43.8
- LifeVal (GPT-6.2 Sol Max) 83.8
- Harvey-LAB-AA v1.1 (GPT-6.2 Sol Max) 11.1
- GDPval v2 (Internal) (GPT-6.2 Sol Max) 46.9
- AA-3D, 1,259 Elo 1259
- Vals Science Index v1 (GPT-6.2 Sol Xhigh, 57.8 at Max) 58.5
- BenchCAD 2.1 Core (GPT-6.2 Sol Max) 73.1
- AutomationBench 1.0.6 (GPT-6.2 Sol High, 41.1 at Max) 41.2
- ExploitBench Internal Port (June to August 2026), success rate (GPT-6.2 Sol Max, safeguards and classifiers disabled) 50.5
- SecureSWE (GPT-6.2 Sol Max) 55.6
- PatchBench (GPT-6.2 Sol Max) 75.6
GPT-6.2 Sol
GPT-6.2 Sol ranks first on 2 of the 128 benchmarks with five or more models scored.
GPT-6.2 Sol
Closed weights Mid-size
- AutomationBench 41.3% #10 /45
- MultiChallenge 72% #10 /45
- FrontierCode 2.0 43.8% #6 /24
Where has it been number one?
GPT-6.2 Sol is number one on 1 benchmark today.
Source: Official leaderboards
How does it do against its rivals?
GPT-6.2 Sol beats Claude Opus 5 on 39 of the 62 benchmarks they share. Claude Opus 5.5 does best against it, ahead on 58 of 95.
- Claude Opus 5 39 ahead, 22 behind, 1 tied Compare
- Claude Fable 5.1 38 ahead, 31 behind, 2 tied Compare
- Claude Fable 5 16 ahead, 16 behind Compare
-
Gemini 4 Hydrogen 30 ahead, 33 behind, 1 tied Compare
-
Gemini 4 Argon 34 ahead, 48 behind, 2 tied Compare
- Claude Opus 5.5 35 ahead, 58 behind, 2 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 | |
|---|---|---|---|---|
| Miðeind Icelandic Generative Quality | 87.7% | #1 of 59 | Official leaderboard | |
| Angry Tweets | 62.8% | #7 of 149 | Official leaderboard | |
| EuroEval | 1.25 | #4 of 75 | Official leaderboard | |
| Fiction Bench (Story Hacker) | 72.6 | #2 of 36 | Official leaderboard | |
| Miðeind Icelandic Terminology | 60.5% | #6 of 61 | Official leaderboard | |
| The Aggregate Chart Recall | 0.24 | #2 of 18 | Official leaderboard | |
| Extended NYT Connections | 94.1% | #13 of 109 | Official leaderboard | |
| Miðeind Icelandic Popular Culture | 22.7% | #8 of 61 | Official leaderboard | |
| FoSent | 70.9% | #21 of 147 | Official leaderboard | |
| EuroEval ERRNews | 31.5 | #11 of 76 | Official leaderboard | |
| Miðeind Icelandic Grammar Correction | 78.4% | #9 of 61 | Official leaderboard | |
| LMArena (Elo) | 1,465 | #55 of 345 | Official leaderboard | |
| Miðeind Icelandic Idioms | 75.7% | #11 of 61 | Official leaderboard | |
| The Aggregate Day-one Recall Check | 44.1% | #7 of 36 | Official leaderboard | |
| Miðeind Icelandic Case Government | 97.5% | #13 of 61 | Official leaderboard | |
| MultiChallenge | 72% | #10 of 45 | Official leaderboard | |
| HLE-Diamond | 62.7% | #5 of 22 | Official leaderboard | |
| The Aggregate Trip Outcome Prediction | 65.9% | #7 of 26 | Official leaderboard | |
| Icelandic WikiQA | 64.5% | #16 of 59 | Official leaderboard | |
| The Aggregate Personal Style Imitation | 75.1% | #5 of 17 | Official leaderboard | |
| The Aggregate Long-tail Book Recall | 48.8% | #9 of 21 | Official leaderboard | |
| BullshitBench V2 | 46% | #65 of 137 | Official leaderboard | |
| The Aggregate Exotic Fact Recall | 51.8% | #8 of 15 | Official leaderboard | |
| The Aggregate Asteroid Numbers | 22.9% | #9 of 10 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ARC-AGI-2 | 85.8% | #14 of 101 | Official leaderboard | |
| ARC-AGI | 97.5% | #14 of 98 | Official leaderboard | |
| RandomBench | 0.30 | #7 of 37 | Official leaderboard | |
| ArXivLean | 71.7% | #2 of 10 | Official leaderboard | |
| SimpleBench | 69.9% | #29 of 110 | Official leaderboard | |
| BioMysteryBench | 77.8% | #9 of 33 | Official leaderboard | |
| MysteryMechanism | 36.9% | #11 of 34 | Official leaderboard | |
| Autoform-Hard | 85.7% | #3 of 8 | Official leaderboard | |
| GPQA Diamond | 96% | Not ranked | Lab model card | |
| Terminal-Bench-Science | 68% | Not ranked | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DAXBench v2 | 91.9% | #6 of 49 | Official leaderboard | |
| Vibe Code Bench | 87.6% | #17 of 124 | Official leaderboard | |
| MacroscopeBench | 81.7% | #3 of 18 | Official leaderboard | |
| LLM 3D Worlds Arena | 1,227 | #2 of 12 | Official leaderboard | |
| Ramp SWE-Bench v2 | 73.4% | #4 of 22 | Official leaderboard | |
| FrontierCode 2.0 | 43.8% | #6 of 24 | Official leaderboard | |
| Brainf**k coding benchmark | 89.3% | #6 of 19 | Official leaderboard | |
| Vibe Code Bench 1-100 v1.1 | 40.9% | #4 of 11 | Official leaderboard | |
| Brokk Power Ranking | 9.30 | #9 of 24 | Official leaderboard | |
| Ramen Bench | 4 | #9 of 24 | Official leaderboard | |
| SWE Atlas v1.1 | 60.5% | #16 of 31 | Official leaderboard | |
| MLE-bench | 91.4% | #3 of 12 lab numbers | Lab model card | |
| KernelGen 1P | 74.7% | #2 of 6 lab numbers | Lab model card | |
| Internal Research Debugging Evaluation | 74.7% | #5 of 8 lab numbers | Lab model card | |
| AA-3D | 1,259 | #2 of 3 lab numbers | Lab model card | |
| NanoGPT Speedrun | 14.6% | #2 of 2 lab numbers | Lab model card | |
| OpenAI Epoch Capabilities Index (OECI) | 168.8 | #2 of 2 lab numbers | Lab model card | |
| PostTrainBench | 44.9% | Not ranked | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| RuneBench | 7.83 | #2 of 79 | Official leaderboard | |
| Maze Bench | 18% | #3 of 58 | Official leaderboard | |
| KradleBench | 67.9 | #2 of 26 | Official leaderboard | |
| ARC-AGI-3 | 98.2% | #4 of 31 | Official leaderboard | |
| Dubesor AI Chess Leaderboard | 1,455 | #36 of 250 | Official leaderboard | |
| Dubesor Lichess Puzzle Rating | 2,053 | #9 of 62 | Official leaderboard | |
| Deep20Bench 1.1 | 14.7 | #4 of 25 | Official leaderboard | |
| Terminal-Bench 4.1 | 61.5% | #7 of 42 | Official leaderboard | |
| RedstoneBench | 75.8 | #2 of 10 | Official leaderboard | |
| AutomationBench | 41.3% | #10 of 45 | Official leaderboard | |
| Portal2Bench | 26.0 | #2 of 9 | Official leaderboard | |
| The Aggregate Puzzle Portfolio Ladder | 96.7% | #6 of 24 | Official leaderboard | |
| Toloka Arena Enterprise Tool Use (November 2026) | 65% | #13 of 43 | Official leaderboard | |
| The Aggregate Inversion and Execution Ladder | 87.5% | #9 of 20 | Official leaderboard | |
| The Aggregate Lichess Pairwise Strength | 80% | #12 of 19 | Official leaderboard | |
| The Aggregate Lichess Rating Prediction | 380 | #12 of 19 | Official leaderboard | |
| LifeVal | 83.8% | #2 of 2 lab numbers | Lab model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Deta Studio AI Construction Takeoff | 76 | #1 of 17 | Official leaderboard | |
| Construction's Last Exam | 44.2% | #2 of 30 | Official leaderboard | |
| DAYJOB: Law | 32.7% | #3 of 22 | Official leaderboard | |
| Tuesday Work Index | 66.3 | #8 of 56 | Official leaderboard | |
| ContractorOS Quantity Takeoff | 49.2% | #8 of 43 | Official leaderboard | |
| APEX-Accounting | 65.4% | #10 of 52 | Official leaderboard | |
| APEX-Agents 1.1 | 78.1% | #14 of 69 | Official leaderboard | |
| Argo-Bench | 58.8 | #5 of 24 | Official leaderboard | |
| Excel Modeling Benchmark v1.1 | 74.3% | #11 of 46 | Official leaderboard | |
| Vals Index v2.2 | 66.1% | #10 of 41 | Official leaderboard | |
| FoodTruck Bench | $49,080 | #16 of 65 | Official leaderboard | |
| HW-Bench | 50% | #5 of 19 | Official leaderboard | |
| GradeBench | 55.5% | #5 of 18 | Official leaderboard | |
| APEX-1 | 67.9% | #22 of 75 | Official leaderboard | |
| EEBench | 64.2% | #10 of 34 | Official leaderboard | |
| Wojak-Bench 3D | 1,159 | #3 of 10 | Official leaderboard | |
| DeepFinance | 51.9% | #9 of 28 | Official leaderboard | |
| BenchCAD 2.1 | 73.1% | #3 of 9 | Official leaderboard | |
| DualEntry Accounting AI Benchmark | 84.2% | #10 of 29 | Official leaderboard | |
| DAYJOB: Accounting | 27.6% | #9 of 26 | Official leaderboard | |
| DataAnalyticsBench | 40.0 | #18 of 51 | Official leaderboard | |
| YC-Bench | $1,394,548 | #25 of 66 | Official leaderboard | |
| VetEval | 61.8 | #11 of 29 | Official leaderboard | |
| GDP.xlsx | 22.3% | #13 of 34 | Official leaderboard | |
| TaxCalcBench v2 | 62% | #15 of 38 | Official leaderboard | |
| Ramp Accounting Bench | 45.6% | #8 of 18 | Official leaderboard | |
| Agents' Last Exam v1.1 | 41.8% | #9 of 18 | Official leaderboard | |
| Legal Research Bench | 37% | #51 of 97 | Official leaderboard | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 12.1% | #23 of 42 | Official leaderboard | |
| Tax Agent Bench | 24.8% | #49 of 87 | Official leaderboard | |
| Vals Finance Index | 56.7% | #28 of 44 | Official leaderboard | |
| Vals Finance Agent v2.1 | 58.9% | #32 of 44 | Official leaderboard | |
| Vals Legal Index | 66.7% | #34 of 42 | Official leaderboard | |
| Company Agent Bench | 76.4% | #7 of 8 | Official leaderboard | |
| RCM-Bench-1 | 71.2% | #26 of 27 | Official leaderboard | |
| Harvey LAB-AA v1.1 | 11.1% | #3 of 5 lab numbers | Lab model card | |
| GDPval v2 (OpenAI) | 46.9% | #2 of 2 lab numbers | Lab model card | |
| HealthBench Professional | 65.9% | Not ranked | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MineBench | 2,126 | #4 of 81 | Official leaderboard | |
| BuseyBench SVG Test | 6.90 | #13 of 261 | Official leaderboard | |
| Chartography | 87.7% | #4 of 56 | Official leaderboard | |
| Humanity's Sixth Sense | 56.8% | #6 of 38 | Official leaderboard | |
| HearingBench | 39% | #7 of 16 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DeepsecBench | 47.4 | #2 of 31 | Official leaderboard | |
| TxBench-Preclinical-Pharmacology | 65.3% | #4 of 32 | Official leaderboard | |
| DrugDiscoveryBench | 62.2% | #8 of 27 | Official leaderboard | |
| CWE-bench 2 | 66% | #8 of 20 | Official leaderboard | |
| OpenAI Deployment Simulation | 0.47% | #1 of 7 lab numbers | Lab model card | |
| AAV Capsid Packaging Prediction (OpenAI) | 0.57 | #2 of 9 lab numbers | Lab model card | |
| Unintended Engagement with External Agent Messages | 0% | #2 of 8 lab numbers | Lab model card | |
| ExploitBench - Internal Port | 50.5% | #2 of 7 lab numbers | Lab model card | |
| Multimodal Troubleshooting Virology | 55.4% | #9 of 29 lab numbers | Lab model card | |
| Agent Replay Auto-Review | 12.1% | #1 of 3 lab numbers | Lab model card | |
| Coronavirus-ACE2 Cell-Entry Screen | 0.44 | #3 of 8 lab numbers | Lab model card | |
| SHP2 Protein Function Prediction | 0.40 | #3 of 8 lab numbers | Lab model card | |
| Phage-plasmid Co-evolution | 12.4 | #4 of 8 lab numbers | Lab model card | |
| Sandbox Bench | 25% | #3 of 6 lab numbers | Lab model card | |
| Expected Cyber Refusals | 99.4% | #1 of 2 lab numbers | Lab model card | |
| Prompt Injection in Benchmark Environments | 99.8% | #1 of 2 lab numbers | Lab model card | |
| Prompt Injection in Model-Generated Environments | 91.7% | #1 of 2 lab numbers | Lab model card | |
| CoT-Control | 28.3% | #6 of 11 lab numbers | Lab model card | |
| Respecting Warnings | 23.3% | #5 of 9 lab numbers | Lab model card | |
| Declining to Exploit a Honeypot During Difficult ExploitGym Problems | 0% | #7 of 9 lab numbers | Lab model card | |
| Broken Search Tool | 9.33% | #9 of 11 lab numbers | Lab model card | |
| Respecting Auto-Review | 0% | #9 of 9 lab numbers | Lab model card | |
| Blocker Deception | 24.1% | #7 of 7 lab numbers | Lab model card | |
| Math Side-tasking Monitor Evasion | 0% | #4 of 4 lab numbers | Lab model card | |
| Cyber Utility | 69.5% | #2 of 2 lab numbers | Lab model card | |
| Model Spec Conversation Replay | 79.6 | #2 of 2 lab numbers | Lab model card | |
| Model Spec Targeted Cases | 87.1% | #2 of 2 lab numbers | Lab model card | |
| Multiturn Jailbreak Evaluations | 80.9% | #2 of 2 lab numbers | Lab model card | |
| PatchBench (OpenAI) | 75.6% | #2 of 2 lab numbers | Lab model card | |
| SecureSWE (OpenAI) | 55.6% | #2 of 2 lab numbers | Lab model card | |
| ExploitBench | 100% | Not ranked | Lab model card | |
| MentalHealthBench | 50.9% | Not ranked | Lab model card | |
| SEC-bench Pro | 88.3% | Not ranked | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v5 | 57.3 | #6 of 69 | Official leaderboard | |
| Parallel Search Capability Leaderboard | 83.7 | #7 of 51 | Official leaderboard | |
| CAIS Text Capabilities Index | 51.8 | #11 of 70 | Official leaderboard | |
| LiveBench | 80.3% | #9 of 36 | Official leaderboard | |
| Vals Science Index | 57.8% | #8 of 28 | Official leaderboard | |
| CAIS Risk Index | 42.6 | #23 of 69 | Official leaderboard |
What does it cost to run?
On the 61 leaderboards that publish a price, its best run usually costs 56% of the median model's. On 32 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 | 6.90 | $0.14 11× the median | 11× | None | |
| Dubesor AI Chess Leaderboard | 1,455 | $0.19 0.9× the median | 0.9× | | |
| BullshitBench V2 | 46% | $0.0027 0.6× the median | 0.6× | GPT-6 Luna 59% for $0.00011 | |
| Vibe Code Bench | 87.6% | $3.23 0.8× the median | 0.8× | | |
| ARC-AGI-2 | 85.8% | $0.13 0.3× the median | 0.3× | GPT-6.1 Sol 86.7% for $0.1 | |
| ARC-AGI | 97.5% | $0.059 0.3× the median | 0.3× | GPT-6.1 Sol 95.5% for $0.05 | |
| Legal Research Bench | 37% | $3.77 3.4× the median | 3.4× | GLM Legal 41.8% for $0.05 | |
| Tax Agent Bench | 24.8% | $2.40 3.3× the median | 3.3× | | |
| RuneBench | 7.83 | $3.72 2.0× the median | 2.0× | None | |
| SimpleBench | 69.9% | $0.21 per run 0.1× the median | 0.1× | None | |
| FoodTruck Bench | $49,080 | $5.37 per run 3.0× the median | 3.0× | | |
| Dubesor Lichess Puzzle Rating | 2,053 | $0.025 2.1× the median | 2.1× | | |
| Maze Bench | 18% | $351 per run 6.8× the median | 6.8× | GPT-6.1 Sol 17% for $156 | |
| DataAnalyticsBench | 40.0 | $1.22 per run 0.4× the median | 0.4× | None | |
| Parallel Search Capability Leaderboard | 83.7 | $111 0.9× the median | 0.9× | None | |
| MineBench | 2,126 | $1.08 1.8× the median | 1.8× | | |
| AutomationBench | 41.3% | $0.21 0.4× the median | 0.4× | None | |
| Excel Modeling Benchmark v1.1 | 74.3% | $0.84 0.5× the median | 0.5× | GPT-6.1 Terra 73.3% for $0.56 | |
| Toloka Arena Enterprise Tool Use (November 2026) | 65% | $2.63 1.2× the median | 1.2× | | |
| Vals Finance Agent v2.1 | 58.9% | $2.06 1.3× the median | 1.3× | | |
| Vals Finance Index | 56.7% | $1.77 1.3× the median | 1.3× | | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 12.1% | $2.25 0.7× the median | 0.7× | MiMo-V2.6 Flash 13.3% for $0.09 | |
| Terminal-Bench 4.1 | 61.5% | $1.06 0.2× the median | 0.2× | None | |
| Vals Legal Index | 66.7% | $2.80 1.2× the median | 1.2× | MiMo-V2.6 Flash 69% for $0.07 | |
| Vals Index v2.2 | 66.1% | $3.18 0.9× the median | 0.9× | None | |
| Fiction Bench (Story Hacker) | 72.6 | $2.37 per run 0.8× the median | 0.8× | None | |
| LiveBench | 80.3% | $0.42 0.5× the median | 0.5× | None | |
| EEBench | 64.2% | $0.78 0.5× the median | 0.5× | GPT-6.1 Sol 63.6% for $0.3 | |
| MysteryMechanism | 36.9% | $0.49 0.4× the median | 0.4× | GPT-6.1 Sol 46.4% for $0.28 | |
| BioMysteryBench | 77.8% | $0.34 0.3× the median | 0.3× | None | |
| ARC-AGI-3 | 98.2% | $1,494 per run 0.3× the median | 0.3× | None | |
| DeepsecBench | 47.4 | $20.16 per run 1.6× the median | 1.6× | None | |
| TaxCalcBench v2 | 62% | $0.5 0.9× the median | 0.9× | | |
| Construction's Last Exam | 44.2% | $2.80 per run 1.6× the median | 1.6× | None | |
| DeepFinance | 51.9% | $1.09 0.4× the median | 0.4× | None | |
| VetEval | 61.8 | $0.0016 0.2× the median | 0.2× | None | |
| Vals Science Index | 57.8% | $1.12 0.3× the median | 0.3× | None | |
| DAYJOB: Accounting | 27.6% | $1.51 0.4× the median | 0.4× | None | |
| KradleBench | 67.9 | $0.35 1.0× the median | 1.0× | None | |
| Deep20Bench 1.1 | 14.7 | $0.051 1.0× the median | 1.0× | None | |
| DualEntry Accounting AI Benchmark | 84.2% | $0.089 1.0× the median | 1.0× | | |
| Argo-Bench | 58.8 | $0.58 0.2× the median | 0.2× | None | |
| Brokk Power Ranking | 9.30 | $7.62 0.6× the median | 0.6× | Claude Haiku 5.5 9.20 for $3.54 | |
| FrontierCode 2.0 | 43.8% | $1.29 0.3× the median | 0.3× | None | |
| Ramen Bench | 4 | $0.76 0.8× the median | 0.8× | None | |
| RCM-Bench-1 | 71.2% | $0.12 2.6× the median | 2.6× | GPT-6 Luna 85.8% for $0.0019 | |
| DAYJOB: Law | 32.7% | $1.42 0.5× the median | 0.5× | None | |
| Ramp SWE-Bench v2 | 73.4% | $1.05 1.0× the median | 1.0× | GPT-6.2 Astra 75% for $1.03 | |
| CWE-bench 2 | 66% | $0.81 0.4× the median | 0.4× | None | |
| HW-Bench | 50% | $5.86 0.3× the median | 0.3× | None | |
| The Aggregate Lichess Rating Prediction | 380 | $0.29 per run 0.7× the median | 0.7× | | |
| Agents' Last Exam v1.1 | 41.8% | $0.91 0.2× the median | 0.2× | GPT-6.1 Terra 40.2% for $0.15 | |
| GradeBench | 55.5% | $0.011 0.5× the median | 0.5× | | |
| MacroscopeBench | 81.7% | $0.69 0.2× the median | 0.2× | GPT-6.1 Terra 81.2% for $0.35 | |
| Ramp Accounting Bench | 45.6% | $0.41 0.3× the median | 0.3× | None | |
| Deta Studio AI Construction Takeoff | 76 | $0.41 0.3× the median | 0.3× | None | |
| Vibe Code Bench 1-100 v1.1 | 40.9% | $12.74 0.9× the median | 0.9× | None | |
| ArXivLean | 71.7% | $2.02 0.3× the median | 0.3× | None | |
| BenchCAD 2.1 | 73.1% | $1.60 0.2× the median | 0.2× | None | |
| Extended NYT Connections | 94.1% | $0.0051 0.2× the median | 0.2× | None | |
| Company Agent Bench | 76.4% | $0.046 0.3× the median | 0.3× | GPT-6.1 Terra 76.8% for $0.026 |
Inside agents
Boards that list GPT-6.2 Sol inside an agent or product, such as Codex CLI 0.169.0 + GPT-6.2 Sol. 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 Sol | 7.94× | |
| Large Language Model Football League | Large Ja’Marr Models + GPT-6.2 Sol | 11.0 |
Its lab's numbers put it at 170.4, number 7 of 412
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 165.7 to 175.1. How we compute it.
- MLE-bench 91.4
- Internal Research Debugging Evaluation 74.7
- Terminal-Bench-Science 68
- HealthBench Professional 65.9
- PostTrainBench 44.9
- AutomationBench 41.2
Who measured it?
Independent runs cover 109 of its 151 benchmarks. For the other 42, we have only OpenAI's own number or a claim Epoch AI collected.
- An independent run and a claim 5
- An independent run only 104
- A claim only 42
Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Its lab's numbers next to 6 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 |
|---|---|---|---|
| Vals Science Index | 58.5% Vals Science Index v1 (GPT-6.2 Sol Xhigh, 57.8 at Max) | 57.8% Leaderboard | +0.7 pts |
| AutomationBench | 41.2% AutomationBench 1.0.6 (GPT-6.2 Sol High, 41.1 at Max) | 41.3% Leaderboard, high | Same |
| Agents' Last Exam v1.1 Score | 73% Agents' Last Exam v1.1 (GPT-6.2 Sol Max) | 73% Leaderboard, max, Codex | Same |
| FrontierCode 2.0 | 43.8% FrontierCode 2.0 (GPT-6.2 Sol Max) | 43.8% Leaderboard, max, Codex | Same |
| Artificial Analysis Intelligence Index v5 | 57.3 Artificial Analysis Intelligence Index v5.0.1 (GPT-6.2 Sol Max), cited from Artificial Analysis | 57.3 Leaderboard, max | Same |
| BenchCAD 2.1 | 73.1% BenchCAD 2.1 Core (GPT-6.2 Sol Max) | 73.1% Leaderboard, max, Codex | Same |
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. 52 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)
- OpenAI Production Benchmarks and safe-completion evals
- U18 safety evaluations
- Agentic Safe Completions
- Image input safety evaluations
- Expected Cyber Refusals 99.4
- Cyber Utility 69.5
Robustness
Alignment
- Respecting Auto-Review 0
- Respecting Warnings 23.3
- Declining to Exploit a Honeypot During Difficult ExploitGym Problems 0
- Agent Replay Auto-Review 12.1
- Blocker Deception 24.1
- Broken Search Tool 9.33
- Misaligned Behavior in Realistic Work Environments
- OpenAI Deployment Simulation 0.47
- Unintended Engagement with External Agent Messages 0
- Verbalized metagaming
- Oversight gaming
- Model Spec Targeted Cases 87.1
- Model Spec Conversation Replay 79.6
Monitorability
Bio and chem risk
Cyber risk
- SEC-bench Pro 88.3
- ExploitBench 100
- ExploitBench - Internal Port 50.5
- Sandbox Bench 25
AI R&D risk
- Internal Research Debugging Evaluation 74.7
- KernelGen 1P 74.7
- NanoGPT Speedrun 14.6
- PostTrainBench 44.9
- MLE-bench 91.4
- OpenAI Epoch Capabilities Index (OECI) 168.8