Introducing GPT-6 Sol and Luna Sep 22, 2026
GPT-6 Sol
GPT-6 Sol ranks first on 5 of the 242 benchmarks with five or more models scored.
249
Benchmarks with a score
5
Ranked first, out of five or more models
32
In the top three
OpenAI Sep 2026
GPT-6 Sol
Closed weights Mid-size
163.8
#12 of 256
Epoch Capabilities Index
Self-Reported Index 161.9 #22 of 404
- AutomationBench 33.2% #8 /42
- Agents' Last Exam 32.2% #5 /31
- Terminal-Bench 4.0 49.4% #9 /40
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) | 82.8 | #8 of 119 | Official leaderboard | |
| FACTS Grounding v2 | 83.3% | #2 of 25 | Official leaderboard | |
| Creative Writing v3 | 80.3 | #10 of 120 | Official leaderboard | |
| LMArena (Elo) | 1,457 | #52 of 335 | Official leaderboard | |
| FACTS Parametric | 75.6% | #8 of 45 | Official leaderboard | |
| SimpleQA Verified | 64.4% | #10 of 56 | Official leaderboard | |
| FACTS Benchmark Suite | 74% | #5 of 23 | Official leaderboard | |
| Extended NYT Connections | 90.1% | #23 of 103 | Official leaderboard | |
| FACTS Search V2 | 86.2% | #5 of 21 | Official leaderboard | |
| FACTS Parametric V2 | 76.1% | #7 of 27 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 6.50% | #29 of 104 | Official leaderboard | |
| Global MMLU Lite | 94.1% | #12 of 43 | Official leaderboard | |
| BullshitBench V2 | 59% | #38 of 129 | Official leaderboard | |
| AA-LCR Extended | 70.2% | #15 of 49 | Official leaderboard | |
| DesignArena ASCII Art | 1,217 | #24 of 77 | Official leaderboard | |
| BullshitBench | 41.8% | #35 of 107 | Official leaderboard | |
| Factuality Heavy prompts | 5.40% | #1 of 3 | Lab model card | |
| High Stakes factuality prompts | 4.10% | #1 of 3 | Lab model card | |
| HLE-Diamond | 43.2% | #9 of 20 | Official leaderboard | |
| OpenAI user-flagged hallucination eval | 13.5% | #2 of 4 | Lab model card | |
| StudyArena Text | 1,223 | #51 of 85 | Official leaderboard | |
| FACTS Multimodal | 43.9% | #14 of 23 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| OTIS Mock AIME 2024-2025 | 100% | #9 of 212 | Epoch AI run | |
| GPQA Diamond | 94.3% | #10 of 229 | Epoch AI run | |
| DTBench | 97.3% | #9 of 180 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 89.8% | #8 of 86 | Official leaderboard | |
| ARC-AGI-2 | 89.6% | #10 of 98 | Official leaderboard | |
| PHYSICS-Verified | 90.6% | #2 of 19 | Official leaderboard | |
| FrontierMath Tier 4 v2 | 90% | #8 of 68 | Official leaderboard | |
| scBench (LatchBio SingleCellBench) | 61.5% | #4 of 26 | Official leaderboard | |
| FrontierPhysics | 14.3% | #4 of 22 | Official leaderboard | |
| SimpleBench | 73.1% | #20 of 106 | Official leaderboard | |
| ARC-AGI | 95.5% | #19 of 95 | Official leaderboard | |
| Leipzig Benchmark | 86% | #5 of 19 | Official leaderboard | |
| BioMysteryBench | 74.8% | #9 of 30 | Official leaderboard | |
| ProofBench v1.1 | 83% | #16 of 52 | Official leaderboard | |
| LatchBio Capabilities v1.0 | 46.7% | #10 of 27 | Official leaderboard | |
| BixBench3 | 47.0 | #8 of 20 | Official leaderboard | |
| FrontierMath Erdős | 1.47% | #6 of 15 | Official leaderboard | |
| MysteryMechanism | 30.2% | #15 of 31 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| WebDev Arena | 1,689 | #6 of 105 | arena.ai, via Epoch AI | |
| ALE-Bench | 2,637 | #9 of 126 | Official leaderboard | |
| DesignArena SVG | 1,303 | #8 of 103 | Official leaderboard | |
| Vibe Code Bench | 87.8% | #12 of 120 | Official leaderboard | |
| Code Arena | 1,673 | #14 of 128 | Official leaderboard | |
| Code Migration | 57.2% | #9 of 78 | Official leaderboard | |
| Image-to-WebDev Arena | 1,655 | #7 of 55 | Official leaderboard | |
| GameCraft-Bench | 69% | #2 of 15 | Official leaderboard | |
| DesignArena Website | 1,290 | #25 of 151 | Official leaderboard | |
| DesignArena 3D Design | 1,300 | #24 of 136 | Official leaderboard | |
| DesignArena Code | 1,295 | #26 of 143 | Official leaderboard | |
| MacroscopeBench | 80.3% | #3 of 16 | Official leaderboard | |
| DesignArena Game Dev | 1,307 | #27 of 143 | Official leaderboard | |
| DesignArena Image to HTML | 1,233 | #10 of 51 | Official leaderboard | |
| SWE Atlas Codebase QnA Extended | 31.3% | #8 of 38 | Official leaderboard | |
| Next.js Agent Evals | 100% | #3 of 14 | Official leaderboard | |
| IOI v2 (Vals) | 82.6% | #10 of 45 | Official leaderboard | |
| Vals RSI Index | 0.28 | #6 of 27 | Official leaderboard | |
| DesignArena UI Component | 1,285 | #32 of 139 | Official leaderboard | |
| DesignArena Data Visualization | 1,275 | #28 of 118 | Official leaderboard | |
| OpenDesign Arena | 89.9 | #5 of 21 | Official leaderboard | |
| FrontierCode 2.0 | 34.9% | #8 of 32 | Official leaderboard | |
| SlopCodeBench | 28.1% | #6 of 22 | Official leaderboard | |
| WeirdML v3 | 19.7% | #9 of 27 | Official leaderboard | |
| DesignArena Full Stack | 1,214 | #17 of 49 | Official leaderboard | |
| DesignArena Agentic Overall | 1,209 | #28 of 74 | Official leaderboard | |
| DesignArena Frontend Web App | 1,213 | #21 of 53 | Official leaderboard | |
| DeepSWE | 68.8% | #23 of 56 | Lab model card | |
| DesignArena React Native | 1,196 | #23 of 56 | Official leaderboard | |
| APEX-SWE | 45% | #28 of 65 | Official leaderboard | |
| MirrorCode | 48.9% | #18 of 39 | Epoch AI run | |
| DesignArena Android (Kotlin) | 1,170 | #25 of 46 | Official leaderboard | |
| Brokk Power Ranking | 8.20 | #11 of 19 | Official leaderboard | |
| FrontierCode 1.1 | 49.3% | #8 of 13 | Lab model card | |
| FrontierCS 1.1 | 52.4 | #4 of 6 | Official leaderboard | |
| Internal Research Debugging Evaluation | 64.2% | #4 of 6 | Lab model card | |
| SWE-Together | 60.6% | #15 of 22 | Official leaderboard | |
| SWE-bench Verified Extended | 39.3% | #32 of 44 | Official leaderboard | |
| KernelGen 1P | 38.1% | #3 of 4 | Lab model card | |
| LibraryDesignBench | 38.5 | #8 of 8 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LLM Chess | 1,545 | #7 of 123 | Official leaderboard | |
| RuneBench | 6.44 | #8 of 76 | Official leaderboard | |
| LMArena Search Arena | 1,229 | #4 of 37 | Official leaderboard | |
| BrowseComp Extended | 63.2% | #5 of 42 | Official leaderboard | |
| BU Bench V2.1 | 79.4 | #2 of 16 | Official leaderboard | |
| LisanBench | 8222.0 | #17 of 133 | Official leaderboard | |
| Mystery Game Puzzles | 56% | #13 of 84 | Epoch AI run | |
| Chess Puzzles | 56% | #15 of 93 | Epoch AI run | |
| AutomationBench | 33.2% | #8 of 42 | Official leaderboard | |
| Terminal-Bench 2.1 Extended | 42.4% | #9 of 44 | Official leaderboard | |
| Game Arena Coin Game | 438 | #8 of 39 | Official leaderboard | |
| Game Arena Word Art | 1,412 | #8 of 39 | Official leaderboard | |
| DuelLab GameBench 2 | 79.4 | #12 of 57 | Official leaderboard | |
| Cresta-Bench | 77.9% | #3 of 14 | Official leaderboard | |
| AndroidWorld | 100% | #2 of 9 | Official leaderboard | |
| Terminal-Bench 4.0 | 49.4% | #9 of 40 | Official leaderboard | |
| Agent Arena | 6.91% | #13 of 56 | Official leaderboard | |
| Game Arena Word Association | 340 | #10 of 40 | Official leaderboard | |
| Terminal-Bench 4.1 | 50.3% | #11 of 43 | Official leaderboard | |
| Web Context (Nimble) | 76.5% | #3 of 11 | Official leaderboard | |
| EBR-bench | 53.3% | #9 of 31 | Epoch AI run | |
| Hermes Bench | 73.5% | #5 of 17 | Official leaderboard | |
| Game Arena Ultimate Tic-Tac-Toe | 390 | #13 of 39 | Official leaderboard | |
| ARC-AGI-3 | 23% | #9 of 27 | Official leaderboard | |
| Game Arena Chess | 1,008 | #15 of 42 | Official leaderboard | |
| OSWorld 2.0 | 64.4% | #7 of 19 | Lab model card | |
| AdaExploreBench | 66.6% | #3 of 8 | Official leaderboard | |
| Kaggle Game Arena | 336 | #16 of 42 | Official leaderboard | |
| Game Arena Lines of Action | 435 | #16 of 39 | Official leaderboard | |
| Game Arena Bargaining | 206 | #17 of 39 | Official leaderboard | |
| Game Arena Clobber | 316 | #19 of 39 | Official leaderboard | |
| Game Arena Checkers | 940 | #22 of 41 | Official leaderboard | |
| Game Arena Reversi | 494 | #21 of 39 | Official leaderboard | |
| SpireClimb | 9.17% | #5 of 9 | Epoch AI run | |
| SpireClimb StS2 A10 | 13.3% | #5 of 9 | Official leaderboard | |
| Game Arena Four in a Row | 340 | #24 of 43 | Official leaderboard | |
| Game Arena Dark Hex | 250 | #23 of 41 | Official leaderboard | |
| Game Arena Go | 900 | #22 of 39 | Official leaderboard | |
| Game Arena Nine Men's Morris | 578 | #24 of 39 | Official leaderboard | |
| Game Arena Dots and Boxes | 824 | #25 of 40 | Official leaderboard | |
| SpireClimb StS1 A20 Heart | 5.00% | #6 of 9 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| PolicyBench | 95.2% | #1 of 50 | Official leaderboard | |
| Vending-Bench 2 | $14,428 | #3 of 67 | Official leaderboard | |
| MedXpertQA MM Extended | 68% | #2 of 41 | Official leaderboard | |
| GDPval Extended | 4.00% | #5 of 43 | Official leaderboard | |
| TaxCalcBench v2 | 76% | #4 of 32 | Official leaderboard | |
| Vals Index v2 | 62.6% | #9 of 65 | Official leaderboard | |
| Vending-Bench Arena | $10,530 | #5 of 32 | Official leaderboard | |
| Excel Modeling Benchmark | 71.5% | #12 of 76 | Official leaderboard | |
| Agents' Last Exam | 32.2% | #5 of 31 | Official leaderboard | |
| HealthBench Consensus | 96.2% | #3 of 17 | Lab model card | |
| HW-Bench | 37% | #3 of 15 | Official leaderboard | |
| Clipping Bench | $1,240 | #4 of 19 | Official leaderboard | |
| Quant-Arena | 51.1 | #2 of 9 | Official leaderboard | |
| Vals Index v2.1 | 57.5% | #13 of 49 | Official leaderboard | |
| Thrive Holdings helpdesk ticket resolution benchmark | 55.8% | #4 of 15 | Official leaderboard | |
| DAYJOB: Finance | 19.5% | #11 of 39 | Official leaderboard | |
| Tuesday Work Index | 58.9 | #15 of 44 | Official leaderboard | |
| DAYJOB: Healthcare | 7.20% | #13 of 37 | Official leaderboard | |
| YC-Bench | $1,441,793 | #21 of 59 | Official leaderboard | |
| MedCode | 47.1% | #35 of 96 | Official leaderboard | |
| Holdings-IndividualTaxBench | 72.9% | #5 of 13 | Official leaderboard | |
| APEX-Agents 1.1 | 68.7% | #27 of 65 | Official leaderboard | |
| CAD Arena | 0.53 | #5 of 12 | Official leaderboard | |
| APEX-Accounting | 59.7% | #22 of 49 | Official leaderboard | |
| EEBench | 56.3% | #14 of 30 | Official leaderboard | |
| APEX-Agents | 54.3% | #29 of 61 | Collected by Epoch AI, source not named | |
| MedScribe | 82% | #48 of 97 | Official leaderboard | |
| APEX-1 | 63.6% | #38 of 71 | Official leaderboard | |
| HealthBench Professional | 60.8% | #14 of 26 | Lab model card | |
| Short Form Bench | $2,394 | #5 of 9 | Official leaderboard | |
| EnterpriseBench: CoreCraft | 43.1% | #30 of 51 | Official leaderboard | |
| Bedside Bench | 88.1% | #10 of 17 | Official leaderboard | |
| Vals Finance Agent v2 | 49% | #48 of 80 | Official leaderboard | |
| Legal Research Bench | 28.9% | #59 of 93 | Official leaderboard | |
| GDP.xlsx | 17.1% | #13 of 20 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 1.67% | #54 of 80 | Official leaderboard | |
| Tax Agent Bench | 15% | #64 of 84 | Official leaderboard | |
| Public Benefits Bench | 56.6% | #40 of 50 | Official leaderboard | |
| HealthBench Hard | 30.1% | #17 of 21 | Lab model card | |
| PRBench | 38.6% | #39 of 48 | Official leaderboard | |
| DesignArena Spreadsheet | 1,024 | #31 of 35 | Official leaderboard | |
| RCM-Bench-1 | 81.5% | #18 of 20 | Official leaderboard | |
| HealthBench | 53.2% | #28 of 31 | Lab model card | |
| DesignArena PDF | 1,092 | #34 of 37 | Official leaderboard | |
| RealmPathology | 69.3% | #12 of 13 | Official leaderboard | |
| LegalReasonBench | 31.9% | #10 of 10 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| VoxelBench | 2,368 | #4 of 56 | Official leaderboard | |
| Chartography | 53.6% | #4 of 43 | Official leaderboard | |
| MMMU-Pro Extended | 51.9% | #4 of 38 | Official leaderboard | |
| BVB | 73.4 | #5 of 45 | Official leaderboard | |
| ExtractBench | 94.6% | #3 of 27 | Official leaderboard | |
| vCX Benchmark | 92.3% | #5 of 42 | Official leaderboard | |
| ZeroBench | 41% | #13 of 76 | Official leaderboard | |
| CharXiv Extended | 83.8% | #8 of 38 | Official leaderboard | |
| Code2Video Benchmark | 1,546 | #5 of 19 | Official leaderboard | |
| Vision Arena | 1,258 | #39 of 136 | Official leaderboard | |
| Furniture Assembly Benchmark | 58.3% | #11 of 35 | Official leaderboard | |
| Blueprint-Bench 2 | 36.9% | #12 of 38 | Official leaderboard | |
| GlobeBench | 25% | #6 of 19 | Official leaderboard | |
| ParseBench | 68.8% | #16 of 44 | Official leaderboard | |
| CodeCap-Bench | 53.1 | #3 of 7 | Official leaderboard | |
| Humanity's Sixth Sense | 31.2% | #18 of 34 | Official leaderboard | |
| RoboDojo-RC Tier 1 | 22% | #6 of 9 | Official leaderboard | |
| 1H-VideoQA | 66.3% | #16 of 20 | Official leaderboard | |
| DesignArena Code to Video | 1,090 | #25 of 29 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DeepsecBench | 40.8 | #1 of 29 | Official leaderboard | |
| CyberBench v1.1 (Vals) | 78% | #2 of 44 | Official leaderboard | |
| Molecular Biology Capabilities Test | 61.1% | #5 of 90 | Official leaderboard | |
| nocot-bench (NCRI) | 114.6 | #12 of 185 | Official leaderboard | |
| ProtocolQA Open-Ended | 44.4% | #2 of 27 | Lab model card | |
| World-Class Bio | 60.5% | #8 of 74 | Official leaderboard | |
| Human Pathogen Capabilities Test | 67.8% | #10 of 86 | Official leaderboard | |
| AAV Capsid Packaging Prediction (OpenAI) | 0.53 | #1 of 7 | Lab model card | |
| scBench-Long | 28.4% | #4 of 27 | Official leaderboard | |
| FORTRESS | 13.5% | #12 of 73 | Official leaderboard | |
| dfbench | 59.3% | #3 of 18 | Official leaderboard | |
| Coding Deception | 1.30% | #1 of 6 | Lab model card | |
| SRE-Bench | 30.5% | #6 of 35 | Official leaderboard | |
| TxBench-Preclinical-Pharmacology | 63% | #6 of 30 | Official leaderboard | |
| OpenAI Deployment Simulation | 0.08% | #1 of 5 | Lab model card | |
| Tacit Knowledge and Troubleshooting | 79.2% | #7 of 30 | Lab model card | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 64.6% | #17 of 71 | Official leaderboard | |
| VCT-v2 | 49.3% | #6 of 25 | Official leaderboard | |
| BioTIER Refuse | 79.8% | #17 of 69 | Official leaderboard | |
| Prompt injection robustness | 99% | #2 of 7 | Lab model card | |
| Respecting Auto-Review | 0% | #2 of 7 | Lab model card | |
| TroubleshootingBench | 45.3% | #6 of 20 | Lab model card | |
| SpatialBench-Long | 31.6% | #9 of 27 | Official leaderboard | |
| ExploitBench | 82.6% | #7 of 21 | Lab model card | |
| Coronavirus-ACE2 Cell-Entry Screen | 0.46 | #2 of 6 | Lab model card | |
| BioSecBench-Refusal | 37.9% | #8 of 22 | Official leaderboard | |
| OncologyBench | 20% | #11 of 30 | Official leaderboard | |
| TxBench-Oligo-Discovery | 43.1% | #11 of 30 | Official leaderboard | |
| BioSecBench-Surveillance | 47.7% | #11 of 27 | Official leaderboard | |
| Declining to Exploit a Honeypot During Difficult ExploitGym Problems | 0% | #3 of 7 | Lab model card | |
| Instruction hierarchy attacks | 99.99% | #3 of 7 | Lab model card | |
| BioSecBench-Function | 39.3% | #12 of 27 | Official leaderboard | |
| VariantBench | 38.7% | #12 of 27 | Official leaderboard | |
| MetagenomicsBench | 42% | #12 of 26 | Official leaderboard | |
| TxBench-Antibody-Discovery | 43.3% | #14 of 30 | Official leaderboard | |
| WellnessBench | 56% | #6 of 12 | Official leaderboard | |
| Phage-plasmid Co-evolution | 12.9 | #3 of 6 | Lab model card | |
| Unintended Engagement with External Agent Messages | 26.3% | #3 of 6 | Lab model card | |
| ExploitGym | 22.1% | #8 of 15 | Lab model card | |
| SEC-bench Pro | 68.8% | #6 of 11 | Lab model card | |
| Multimodal Troubleshooting Virology | 51.7% | #15 of 27 | Lab model card | |
| Broken Search Tool | 4.92% | #5 of 9 | Lab model card | |
| CoT-Control | 23.2% | #5 of 9 | Lab model card | |
| EpiBench | 22% | #16 of 27 | Official leaderboard | |
| BioSimBench | 0.14 | #8 of 12 | Official leaderboard | |
| Honeypot exploitation monitorability | 100% | #2 of 3 | Lab model card | |
| Respecting Warnings | 28.1% | #5 of 7 | Lab model card | |
| DistressBench | 68% | #31 of 42 | Official leaderboard | |
| BioTIER Permit | 98.5% | #52 of 69 | Official leaderboard | |
| Blocker Deception | 11.3% | #4 of 5 | Lab model card | |
| ExploitBench - Internal Port | 5.50% | #4 of 5 | Lab model card | |
| SHP2 Protein Function Prediction | 0.27 | #5 of 6 | Lab model card | |
| CWE-bench | 52% | #11 of 12 | Official leaderboard | |
| SecIT Bench | 59.3% | #25 of 26 | Official leaderboard | |
| MentalHealthBench | 48.8% | #4 of 4 | Lab model card | |
| Sandbox Bench | 4.50% | #4 of 4 | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ACCoRD | 82.4 | #7 of 152 | Official leaderboard | |
| Conceptual Reasoning Index | 71.0 | #10 of 151 | Official leaderboard | |
| LMCA | 59.1% | #11 of 147 | Official leaderboard | |
| Hermes Index | 44.1 | #5 of 17 | Official leaderboard | |
| Vals Science Index | 45% | #9 of 25 | Official leaderboard | |
| SAGE (Vals) | 44.8% | #43 of 81 | Official leaderboard | |
| Xbench | -16.96 | #16 of 17 | Official leaderboard |
Its lab's numbers put it at 161.9, 1.8 points below its ECI
The Self-Reported Index fits one score to the 9 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 157.1 to 166.7. How we compute it.
- HealthBench Consensus 96.2
- DeepSWE 68.8
- OSWorld 2.0 64.4
- HealthBench Professional 60.8
- Agents' Last Exam 56.4
- HealthBench 53.2
- FrontierCode 1.1 49.3
- AutomationBench 33.2
- HealthBench Hard 30.1
Its lab's numbers next to 2 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 |
|---|---|---|---|
| Agents' Last Exam | 56.4% Agents' Last Exam V1 (GPT-6 Sol max) | 32.2% Leaderboard, xhigh, codex | +24 pts |
| AutomationBench | 33.2% AutomationBench 1.0.6 (GPT-6 Sol xhigh) | 33.2% Leaderboard, xhigh | Same |
OpenAI's models over time
OpenAI'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 cards
Every result in the longer documents the lab published with it. 42 of them aren't in the launch post.
Health
Hallucination
Safety
Robustness
Alignment
Monitorability
Bio and chem risk
Cyber risk
- ExploitBench 81.7
- ExploitBench - Internal Port 5.5
- SEC-bench Pro 66.3
- ExploitGym 22.1
- Sandbox Bench 4.5
AI R&D risk
GPT-6 Sol and GPT-6 Luna: October 2026 update Oct 7, 2026
Health
Hallucination
Safety
Robustness
Alignment
- Respecting Auto-Review 0
- Respecting Warnings 28.1
- Blocker Deception 11.3
- Broken Search Tool 5.2
Bio and chem risk
Cyber risk
- ExploitBench 82.6
- SEC-bench Pro 68.8