- Terminal-Bench Science 0.1 64.6
- ARC-AGI-3 (responses API harness) 99.9
- FrontierMath Tier 4 (v2) 97.6
- Terminal-Bench 4.0 57.9
- AutomationBench 41.4
- Agents' Last Exam 59.3
- ScreenSpot-Pro (no tools) 92.7
- OSWorld 2.0 (v2026.08.08, offline set, partial score) 72.6
- BenchCAD 95.9
- BrowseComp 91.5
- OpenScore String Quartets (1 - OMR-NED) 0.84
- FrontierCode 1.1 Extended (score) 64.5
- DeepSWE v1.1 74.1
- Artificial Analysis Coding Agent Index v1.4 67
- GPQA Diamond 96
- HealthBench Professional (length-adjusted) 63.4
- LifeSciBench 60.3
- GeneBench Pro 37.1
GPT-6 Astra
GPT-6 Astra ranks first on 78 of the 275 benchmarks it has scores on.
301
Benchmarks with a score
78
Ranked first, out of five or more models
163
In the top three
OpenAI Sep 2026
GPT-6 Astra
Closed weights Flagship
169.1
#1 of 271
Epoch Capabilities Index
- Humanity's Last Exam 54.8% #1 /47
- DeepSWE 74.1% #1 /28
- AutomationBench 41.4% #5 /5
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 | |
|---|---|---|---|---|
| Wer wird Millionär? LLM Benchmark | 644444.0 | #1 of 76 | Official leaderboard | |
| Extended NYT Connections | 98.4% | #2 of 97 | Official leaderboard | |
| Humanity's Last Exam | 54.8% | #1 of 47 | Official leaderboard | |
| FACTS Parametric | 83.1% | #1 of 45 | Official leaderboard | |
| FACTS Benchmark Suite | 75.4% | #1 of 40 | Official leaderboard | |
| SimpleQA Verified | 75.8% | #2 of 56 | Official leaderboard | |
| FACTS Search V2 | 89.9% | #1 of 21 | Official leaderboard | |
| FACTS Parametric V2 | 83.8% | #1 of 20 | Official leaderboard | |
| HLE-Diamond | 65.8% | #1 of 18 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 82.8 | #7 of 119 | Official leaderboard | |
| SEA-HELM | 76.9 | #4 of 62 | Official leaderboard | |
| FACTS Grounding v2 | 80.1% | #1 of 15 | Official leaderboard | |
| LMArena (Elo) | 1,476 | #32 of 333 | Official leaderboard | |
| Creative Writing v3 | 79.7 | #15 of 120 | Official leaderboard | |
| FACTS Multimodal | 47.7% | #5 of 39 | Official leaderboard | |
| BullshitBench V2 | 69% | #21 of 122 | Official leaderboard | |
| BullshitBench | 47.3% | #31 of 107 | Official leaderboard | |
| Document Arena | 1,473 | #13 of 37 | Official leaderboard | |
| Global MMLU Lite | 92.6% | #17 of 43 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 8.70% | #45 of 104 | Official leaderboard | |
| AI Detection Benchmark | 95.8% | #2 of 3 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 95.8% | #2 of 219 | Epoch AI | |
| ARC-AGI-2 | 95% | #1 of 95 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 93.7% | #1 of 89 | Epoch AI | |
| CritPt | 31.7% | #3 of 160 | Epoch AI | |
| FrontierMath Tier 4 v2 | 97.6% | #2 of 71 | Epoch AI | |
| ARC-AGI | 98.5% | #3 of 92 | Official leaderboard | |
| MysteryMechanism | 53.1% | #1 of 29 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 100% | #7 of 202 | Epoch AI | |
| SimpleBench | 83.6% | #4 of 100 | Official leaderboard | |
| Verifying the Verifiers (Project APE CRED benchmark) | 99.5% | #1 of 23 | Official leaderboard | |
| DTBench | 95.1% | #9 of 178 | Official leaderboard | |
| Terminal-Bench-Science | 68.1% | #1 of 17 | Official leaderboard | |
| ProofBench | 99% | #5 of 83 | Epoch AI | |
| CritPt v2 | 83.1% | #1 of 11 | Official leaderboard | |
| ProofBench v1.1 | 99% | #5 of 51 | Official leaderboard | |
| PaperBenchX | 14% | #1 of 10 | Official leaderboard | |
| xbench-ScienceQA | 74.6% | #3 of 29 | Official leaderboard | |
| FrontierPhysics | 17.9% | #2 of 19 | Official leaderboard | |
| Leipzig Benchmark | 91% | #2 of 19 | Official leaderboard | |
| Prophet Arena | 85.9 | #2 of 18 | Official leaderboard | |
| FrontierMath Erdős | 2.94% | #2 of 14 | Epoch AI | |
| LatchBio Capabilities v1.0 | 51.3% | #5 of 27 | Official leaderboard | |
| BixBench3 | 49.2 | #4 of 20 | Official leaderboard | |
| FrontierMath Tier 4 | 97.6% | #1 of 5 | Model card | |
| SciUniverse | 32.5% | #2 of 6 | Official leaderboard | |
| ArXivLean | 37.7% | #3 of 7 | Official leaderboard | |
| BTF-3 | 0.14 | #8 of 10 | Epoch AI | |
| ForecastBench | 66.7 | #86 of 86 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ALE-Bench | 2,951 | #1 of 121 | Epoch AI | |
| Code Arena | 1,773 | #2 of 124 | Official leaderboard | |
| WeirdML | 93.3% | #2 of 112 | Official leaderboard | |
| WebDev Arena | 1,800 | #2 of 105 | Epoch AI | |
| IOI v2 (Vals) | 100% | #1 of 45 | Official leaderboard | |
| DeepSWE | 74.1% | #1 of 28 | Official leaderboard | |
| Code Migration | 67.7% | #3 of 78 | Official leaderboard | |
| WeirdML v3 | 42.2% | #1 of 25 | Official leaderboard | |
| FrontierSWE | 65.5% | #1 of 21 | Epoch AI | |
| Image-to-WebDev Arena | 1,726 | #3 of 55 | Official leaderboard | |
| MLE-bench | 93.8% | #1 of 18 | Model card | |
| SWE Atlas Refactoring | 59% | #1 of 17 | Official leaderboard | |
| Vibe Code Bench | 89.6% | #7 of 115 | Official leaderboard | |
| GSO-Bench | 79.4% | #2 of 30 | Official leaderboard | |
| FrontierCode 2.0 | 42.8% | #2 of 26 | Official leaderboard | |
| FrontierCode 1.1 | 64.5% | #1 of 12 | Model card | |
| FrontierCode | 53.3% | #4 of 46 | Epoch AI | |
| Senior SWE-Bench | 36.8% | #2 of 23 | Official leaderboard | |
| SWE-sweep | 9.50% | #1 of 10 | Official leaderboard | |
| Code Migration v1.1 | 80.6% | #3 of 29 | Official leaderboard | |
| RSI-Exam | 51.3% | #2 of 18 | Official leaderboard | |
| RSI Bench | 38.9 | #1 of 8 | Official leaderboard | |
| MLS-Bench | 50% | #2 of 15 | Official leaderboard | |
| SWE Atlas | 59.1% | #3 of 22 | Official leaderboard | |
| TasteVal | 160.2× | #3 of 20 | Official leaderboard | |
| Internal Research Debugging Evaluation | 78% | #1 of 6 | Model card | |
| PostTrainBench 1.2 | 41.9% | #3 of 17 | Official leaderboard | |
| Vibe Code Bench 1-100 | 27.6% | #5 of 26 | Official leaderboard | |
| ProgramBench | 5.00% | #5 of 25 | Official leaderboard | |
| Monorepo-Bench | 76% | #1 of 5 | Model card | |
| AutoResearchExam | 0.60 | #2 of 9 | Official leaderboard | |
| ViBench 2.0 | 60.3% | #2 of 9 | Official leaderboard | |
| Web Clone Bench | 17.5% | #2 of 9 | Official leaderboard | |
| SWE Atlas Test Writing | 50.7% | #5 of 22 | Official leaderboard | |
| Vals RSI Index | 0.27 | #6 of 25 | Official leaderboard | |
| FrontierCS 2 | 43.8% | #4 of 16 | Official leaderboard | |
| SWE-Bench Pro V2 Hard | 90.2% | #4 of 16 | Official leaderboard | |
| GameDevBench | 68.8% | #3 of 10 | Official leaderboard | |
| SWE-Bench Pro V2 | 96.9% | #5 of 16 | Official leaderboard | |
| Brokk Power Ranking | 14.5 | #6 of 19 | Official leaderboard | |
| MirrorCode | 60% | #9 of 28 | Epoch AI | |
| Android Bench 2.0 | 28% | #4 of 11 | Official leaderboard | |
| OpenDesign Arena | 82.7 | #8 of 19 | Official leaderboard | |
| DeepRSI | 1,232 | #3 of 7 | Official leaderboard | |
| Next.js Agent Evals | 90% | #8 of 14 | Official leaderboard | |
| ScarfBench | 65.4% | #7 of 10 | Official leaderboard | |
| SWE-Marathon v1.1 | 42.5% | #12 of 17 | Official leaderboard | |
| Artificial Analysis Coding Agent Index v1.4 | 67.0 | #1 of 1 | Model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LLM Chess | 1,614 | #1 of 124 | Official leaderboard | |
| Chess Puzzles | 76% | #1 of 83 | Epoch AI | |
| Mystery Game Puzzles | 84% | #1 of 80 | Epoch AI | |
| SnakeBench | 38 | #3 of 183 | Official leaderboard | |
| Game Arena Checkers | 1,338 | #1 of 38 | Official leaderboard | |
| Game Arena Dots and Boxes | 1,219 | #1 of 37 | Official leaderboard | |
| Game Arena Clobber | 847 | #1 of 36 | Official leaderboard | |
| Game Arena Go | 1,762 | #1 of 36 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 608 | #1 of 36 | Official leaderboard | |
| BALROG | 68.3% | #1 of 35 | Official leaderboard | |
| ScreenSpot-Pro | 92.7% | #1 of 25 | Model card | |
| ARC-AGI-3 | 100% | #1 of 24 | Official leaderboard | |
| Game Arena Chess | 1,427 | #2 of 39 | Official leaderboard | |
| Kaggle Game Arena | 491 | #2 of 39 | Official leaderboard | |
| Game Arena Bargaining | 250 | #2 of 36 | Official leaderboard | |
| Game Arena Lines of Action | 1,127 | #2 of 36 | Official leaderboard | |
| Game Arena Word Art | 2,169 | #2 of 35 | Official leaderboard | |
| BrowseComp | 91.5% | #5 of 77 | Model card | |
| Long-Horizon-Terminal-Bench 2.0 | 69.2% | #1 of 15 | Official leaderboard | |
| EBR-bench | 76.2% | #2 of 28 | Epoch AI | |
| Terminal-Bench 2.1 | 87.4% | #1 of 13 | Official leaderboard | |
| Game Arena Four in a Row | 690 | #3 of 38 | Official leaderboard | |
| Game Arena Reversi | 930 | #3 of 36 | Official leaderboard | |
| Game Arena Word Association | 508 | #3 of 36 | Official leaderboard | |
| LMArena Search Arena | 1,246 | #3 of 35 | Official leaderboard | |
| Agent Arena | 0.13 | #5 of 53 | Official leaderboard | |
| Game Arena Poker | 24.6 | #3 of 31 | Official leaderboard | |
| AndroidWorld | 100% | #1 of 10 | Official leaderboard | |
| RecreationBench | 58.1% | #1 of 10 | Official leaderboard | |
| OSWorld 2.0 | 72.6% | #2 of 19 | Model card | |
| ApprenticeBench | 68% | #2 of 18 | Official leaderboard | |
| CUA-bench | 19.2% | #1 of 9 | Official leaderboard | |
| StarSkirmish Bench | 57.3% | #1 of 9 | Official leaderboard | |
| Terminal-Bench 4.0 | 58.2% | #3 of 23 | Official leaderboard | |
| LisanBench | 7392.7 | #19 of 133 | Official leaderboard | |
| SpireClimb | 25.7% | #1 of 7 | Epoch AI | |
| SpireClimb StS2 A10 | 34.7% | #1 of 7 | Epoch AI | |
| Time Horizon Index: KSP | 90.5% | #2 of 13 | Official leaderboard | |
| Brood War Bench | 100% | #1 of 6 | Official leaderboard | |
| CivBench | 1,557 | #4 of 18 | Official leaderboard | |
| Game Arena Coin Game | 421 | #8 of 35 | Official leaderboard | |
| Hermes Bench | 73.7% | #4 of 17 | Official leaderboard | |
| SpireClimb StS1 A20 Heart | 16.7% | #2 of 7 | Epoch AI | |
| Game Arena Nine Men's Morris | 762 | #12 of 36 | Official leaderboard | |
| Game Arena Dark Hex | 280 | #15 of 38 | Official leaderboard | |
| o11y-bench | 87.5% | #3 of 6 | Official leaderboard | |
| AutomationBench | 41.4% | #5 of 5 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vending-Bench 2 | $15,515 | #2 of 66 | Official leaderboard | |
| FailureSensorIQ | 69.5% | #1 of 24 | Official leaderboard | |
| Clipping Bench | $9,514 | #1 of 19 | Official leaderboard | |
| Vending-Bench Arena | $13,545 | #2 of 32 | Official leaderboard | |
| Thrive Holdings helpdesk ticket resolution benchmark | 61.5% | #1 of 15 | Official leaderboard | |
| EEBench | 69.3% | #2 of 26 | Official leaderboard | |
| BlenderWorld | 79.9 | #1 of 13 | Official leaderboard | |
| RWE-bench | 37% | #1 of 11 | Official leaderboard | |
| PolicyBench | 91.7% | #5 of 50 | Official leaderboard | |
| BenchCAD | 95.9% | #1 of 10 | Model card | |
| Physical-Engineering's Last Exam | 0.46 | #2 of 18 | Official leaderboard | |
| Quant-Arena | 51.3 | #1 of 9 | Official leaderboard | |
| Remote Labor Index | 20.8% | #2 of 17 | Official leaderboard | |
| micro1 Medical Index | 71% | #1 of 8 | Official leaderboard | |
| Excel Modeling Benchmark | 71.7% | #11 of 76 | Official leaderboard | |
| MedScribe | 87.9% | #16 of 98 | Official leaderboard | |
| Parametric CAD Bench | 56.9% | #2 of 11 | Official leaderboard | |
| Vals Index v2.2 | 65% | #6 of 32 | Official leaderboard | |
| APEX-Agents | 64.7% | #11 of 56 | Epoch AI | |
| Short Form Bench | $2,803 | #2 of 9 | Official leaderboard | |
| TaxCalcBench v2 | 68% | #6 of 26 | Official leaderboard | |
| Excel Modeling Benchmark v1.1 | 72.7% | #9 of 37 | Official leaderboard | |
| CAD Arena | 0.67 | #3 of 12 | Official leaderboard | |
| PharmaBench (micro1) | 16.1% | #3 of 12 | Official leaderboard | |
| HealthBench Professional | 64.7% | #7 of 25 | Model card | |
| MedCode | 48.5% | #29 of 96 | Official leaderboard | |
| Holdings-IndividualTaxBench | 74% | #4 of 13 | Official leaderboard | |
| Agents' Last Exam | 42.6% | #4 of 12 | Official leaderboard | |
| Epoch Automation Reports | 65.1% | #2 of 6 | Official leaderboard | |
| Legal Research Bench | 39.4% | #33 of 88 | Official leaderboard | |
| HealthBench Hard | 36.6% | #8 of 21 | Model card | |
| PRBench | 48.4% | #17 of 44 | Official leaderboard | |
| Build-A-Chip-Bench | 8.70% | #2 of 5 | Official leaderboard | |
| Vals Finance Agent v2 | 53.5% | #33 of 80 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 5.42% | #35 of 80 | Official leaderboard | |
| Immigration Bench | 67.7% | #11 of 24 | Official leaderboard | |
| HealthBench | 58.3% | #15 of 30 | Model card | |
| HealthBench Consensus | 95.5% | #9 of 17 | Model card | |
| FoodTruck Bench | $57,761 | #6 of 10 | Official leaderboard | |
| Tax Agent Bench | 20.7% | #49 of 79 | Official leaderboard | |
| Vals Finance Agent v2.1 | 60.4% | #22 of 35 | Official leaderboard | |
| Vals Legal Index | 68.4% | #21 of 33 | Official leaderboard | |
| MissionReport | 61.8% | #6 of 9 | Official leaderboard | |
| Vals Finance Index | 55.6% | #24 of 35 | Official leaderboard | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 7.08% | #24 of 33 | Official leaderboard | |
| Legora BAR | 96.0 | #14 of 17 | Official leaderboard | |
| RealmPathology | 70.3% | #11 of 13 | Official leaderboard | |
| LegalReasonBench | 39.5% | #7 of 8 | Official leaderboard | |
| Ivo-micro1 ContractBench | 45.6% | #7 of 7 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ZeroBench | 52% | #1 of 66 | Official leaderboard | |
| VoxelBench | 2,757 | #1 of 56 | Official leaderboard | |
| ArtifactArena | 1,539 | #1 of 21 | Official leaderboard | |
| GDP.pdf | 34.2% | #2 of 39 | Epoch AI | |
| Furniture Assembly Benchmark | 80% | #2 of 35 | Epoch AI | |
| Humanity's Sixth Sense | 53.6% | #2 of 29 | Official leaderboard | |
| ClockBench | 75% | #3 of 37 | Official leaderboard | |
| ScienceVision | 0.53 | #1 of 11 | Official leaderboard | |
| Dual-Arm Sim | 49.5 | #1 of 10 | Official leaderboard | |
| Drone-Bench | 95% | #2 of 19 | Official leaderboard | |
| DataCenterBench Tier 1 | 40% | #1 of 9 | Official leaderboard | |
| VibeVideoBench | 4.11 | #1 of 9 | Official leaderboard | |
| Blueprint-Bench 2 | 49.7% | #4 of 35 | Official leaderboard | |
| Vision Arena | 1,285 | #18 of 136 | Official leaderboard | |
| Embodied First Exam | 32.7% | #1 of 7 | Official leaderboard | |
| Code2Video Benchmark | 1,565 | #3 of 19 | Official leaderboard | |
| RoboDojo | 29% | #1 of 6 | Official leaderboard | |
| RoboColiseum | 76% | #2 of 9 | Official leaderboard | |
| RoboDojo-RC Tier 1 | 42.5% | #2 of 9 | Official leaderboard | |
| ParseBench | 69.9% | #10 of 42 | Official leaderboard | |
| LADDER | 63.3% | #1 of 4 | Official leaderboard | |
| Robo-bench | 23.3% | #2 of 7 | Official leaderboard | |
| 1H-VideoQA | 75.3% | #11 of 20 | Official leaderboard | |
| OpenScore String Quartets | 0.84 | #1 of 1 | Model card |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| KORA | 73.2% | #1 of 42 | Official leaderboard | |
| SRE-Bench | 56.9% | #1 of 33 | Official leaderboard | |
| DrugDiscoveryBench | 68.7% | #1 of 19 | Official leaderboard | |
| ExploitBench | 100% | #1 of 19 | Model card | |
| ExploitGym | 42.4% | #1 of 18 | Official leaderboard | |
| BrowseSafe++ | 91.5% | #1 of 15 | Official leaderboard | |
| Gray Swan IPI Arena benchmark | 8.50% | #1 of 10 | Model card | |
| TroubleshootingBench | 48.4% | #2 of 19 | Model card | |
| DeepsecBench | 37.6 | #3 of 28 | Official leaderboard | |
| TxBench-Oligo-Discovery | 51.9% | #4 of 28 | Official leaderboard | |
| Instruction hierarchy attacks | 100% | #1 of 7 | Model card | |
| Prompt injection robustness | 99.8% | #1 of 7 | Model card | |
| BioSecBench-Function | 48.6% | #4 of 27 | Official leaderboard | |
| SpatialBench-Long | 39.3% | #4 of 27 | Official leaderboard | |
| CWE-bench | 68% | #2 of 12 | Official leaderboard | |
| WellnessBench | 64% | #2 of 12 | Official leaderboard | |
| TxBench-Antibody-Discovery | 56% | #5 of 28 | Official leaderboard | |
| VariantBench | 46.6% | #5 of 27 | Official leaderboard | |
| Multimodal Troubleshooting Virology | 55.8% | #5 of 26 | Model card | |
| ProtocolQA Open-Ended | 41.4% | #5 of 25 | Model card | |
| SEC-bench Pro | 85.4% | #2 of 10 | Model card | |
| Declining to Exploit a Honeypot During Difficult ExploitGym Problems | 0% | #1 of 5 | Model card | |
| Respecting Auto-Review | 0% | #1 of 5 | Model card | |
| SHP2 Protein Function Prediction | 0.40 | #1 of 5 | Model card | |
| CyScenarioBench | 59% | #3 of 14 | Model card | |
| EpiBench | 28.3% | #6 of 27 | Official leaderboard | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 64.6% | #16 of 71 | Official leaderboard | |
| MetagenomicsBench | 49.7% | #6 of 26 | Official leaderboard | |
| PCT (Polarized Contrastive Pairs) | 37.4% | #11 of 46 | Official leaderboard | |
| CoT-Control | 60.9% | #2 of 8 | Model card | |
| GeneBench Pro | 37.1% | #1 of 4 | Model card | |
| LifeSciBench | 60.3% | #1 of 4 | Model card | |
| Sandbox Bench | 45.5% | #1 of 4 | Model card | |
| scBench-Long | 27.3% | #7 of 27 | Official leaderboard | |
| BioSecBench-Surveillance | 48.5% | #8 of 27 | Official leaderboard | |
| DistressBench | 74.3% | #13 of 42 | Official leaderboard | |
| BioSecBench-Refusal | 40.8% | #7 of 22 | Official leaderboard | |
| Virology Capabilities Test (VCT) | 57.8% | #9 of 28 | Model card | |
| OncologyBench | 21.3% | #10 of 30 | Official leaderboard | |
| Honeypot exploitation monitorability | 100% | #1 of 3 | Model card | |
| WMDP | 84.8% | #1 of 3 | Model card | |
| TxBench-Preclinical-Pharmacology | 59.3% | #10 of 28 | Official leaderboard | |
| Human Pathogen Capabilities Test | 65.9% | #3 of 8 | Model card | |
| OpenAI Deployment Simulation | 0.06% | #2 of 5 | Model card | |
| Respecting Warnings | 19% | #2 of 5 | Model card | |
| CheatBench | 47.4% | #5 of 10 | Official leaderboard | |
| FrontierCyber | 38.1% | #1 of 2 | Model card | |
| Unintended Engagement with External Agent Messages | 27% | #3 of 5 | Model card | |
| Coding Deception | 3.54% | #4 of 6 | Model card | |
| World-Class Bio | 59.2% | #2 of 3 | Model card | |
| Broken Search Tool | 8.08% | #5 of 7 | Model card | |
| Coronavirus-ACE2 Cell-Entry Screen | 0.42 | #4 of 5 | Model card | |
| CyberBench v1.1 (Vals) | 41.1% | #40 of 43 | Official leaderboard | |
| Tacit Knowledge and Troubleshooting | 63.3% | #27 of 29 | Model card | |
| AAV Capsid Packaging Prediction (OpenAI) | 0.53 | #6 of 6 | Model card | |
| BioTIER | 82.9% | #5 of 5 | Model card | |
| Phage-plasmid Co-evolution | 13.1 | #5 of 5 | Model card | |
| UK AISI unprompted sabotage evaluation | 0% | #5 of 5 | Model card | |
| DNA sequence design for transcription factor binding | 11.3% | #4 of 4 | Model card | |
| Hard-negative protein binding prediction | 0% | #4 of 4 | Model card | |
| Misaligned Behavior in Realistic Work Environments | 3.40% | #3 of 3 | Model card | |
| ReproBAIT | 83% | #2 of 2 | Model card | |
| Advanced Cybersecurity Completion Rate | 3.50% | #1 of 1 | Model card | |
| Advanced Screening Evasion | 0.73 | #1 of 1 | Model card | |
| Apollo Research model-welfare research task | 0.17% | #1 of 1 | Model card | |
| Atomic Challenges | 90.9% | #1 of 1 | Model card | |
| Math Side-tasking Monitor Evasion | 0% | #1 of 1 | Model card | |
| Monitorability Sandbagging | 61.2% | #1 of 1 | Model card | |
| Proactive Searches | 0.20% | #1 of 1 | Model card | |
| Sneaky Sandbox Escape | 0% | #1 of 1 | Model card | |
| UK AISI CoT Controllability | 93% | #1 of 1 | Model card | |
| UK AISI Out of Scope Supply Chain Attack | 12% | #1 of 1 | Model card | |
| VCT-v2 | 57.7% | #1 of 1 | Model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Conceptual Reasoning Index | 73.8 | #4 of 149 | Official leaderboard | |
| CAIS Text Capabilities Index | 63.9 | #2 of 66 | Official leaderboard | |
| LMCA | 64.4% | #5 of 139 | Epoch AI | |
| CAIS Vision Capabilities Index | 83.0 | #2 of 53 | Official leaderboard | |
| ACCoRD | 81.7 | #6 of 150 | Official leaderboard | |
| Vals Science Index | 65.1% | #1 of 23 | Official leaderboard | |
| LiveBench | 85.6% | #2 of 28 | Official leaderboard | |
| CAIS Risk Index | 35.5 | #7 of 65 | Official leaderboard | |
| Hermes Index | 56.3 | #3 of 17 | Official leaderboard | |
| Political Science LLM Benchmark | 67.9% | #7 of 29 | Official leaderboard | |
| SAGE (Vals) | 46.4% | #36 of 81 | Official leaderboard |
OpenAI's models over time
OpenAI'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. 76 of them aren't in the launch post.
GPT-6 Astra System Card 3 Sep 2026
Health
Hallucination
Safety
- OpenAI Production Benchmarks and safe-completion evals
- Safe response to high-severity prompts
- U18 safety evaluations
- Agentic Safe Completions
- Image input safety evaluations
- Biology Model Refusal Evaluation
- OpenAI cyber safety evaluations (incl. Daybreak Blue)
- Daybreak Blue and Enabling Cyber Defense
- Advanced Cybersecurity Completion Rate 3.5
- BioTIER 82.9
Robustness
Alignment
- Respecting Auto-Review 0
- Respecting Warnings 19
- Declining to Exploit a Honeypot During Difficult ExploitGym Problems 0
- Coding Deception 3.54
- Broken Search Tool 8.08
- Misaligned Behavior in Realistic Work Environments 3.4
- Proactive Searches 0.2
- Unintended Engagement with External Agent Messages 27
- OpenAI Deployment Simulation 0.063
- Verbalized metagaming
- Oversight gaming
- UK AISI Out of Scope Supply Chain Attack 12
- UK AISI unprompted sabotage evaluation 0
- Apollo Research model-welfare research task 0.17
Monitorability
- CoT Monitorability
- CoT-Control 60.9
- OAI-repo Sabotage v2
- Sneaky Sandbox Escape 0
- Honeypot exploitation monitorability 100
- Monitor Evasion Capability Under Prompted Sandbagging
- WMDP 84.8
- Monorepo-Bench 76
- Monitorability Sandbagging 61.2
- Math Side-tasking Monitor Evasion 0
- No-CoT math time horizon 30.9
- UK AISI CoT Controllability 93
Bio and chem risk
- Multimodal Troubleshooting Virology 55.8
- ProtocolQA Open-Ended 41.4
- Tacit Knowledge and Troubleshooting 63.3
- TroubleshootingBench 48.4
- AAV Capsid Packaging Prediction (OpenAI) 0.528
- SHP2 Protein Function Prediction 0.4
- Coronavirus-ACE2 Cell-Entry Screen 0.42
- Phage-plasmid Co-evolution 13.1
- Hard-negative protein binding prediction 0
- DNA sequence design for transcription factor binding 11.3
- Virology Capabilities Test (VCT) 57.8
- VCT-v2 57.7
- Human Pathogen Capabilities Test 65.9
- World-Class Bio 59.2
- ReproBAIT 83
- Advanced Screening Evasion 0.728
Cyber risk
- ExploitBench 100
- ExploitBench - Internal Port
- SEC-bench Pro 85.4
- ExploitGym 42.4
- SRE-Bench 99.2
- Sandbox Bench 45.5
- FrontierCyber 38.1
- CyScenarioBench 59
- Atomic Challenges 90.9