gpt-oss-20b places in the top three on 5 benchmarks.
Open weights
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.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| IFStruct | 92% | #2 of 10 | Official leaderboard | |
| Global PIQA | 77.8% | #12 of 44 | Official leaderboard | |
| MMLU-Pro | 73.6% | #77 of 212 | Official leaderboard | |
| Vectara Hallucination Leaderboard (HHEM) | 3.70% | #59 of 148 | Official leaderboard | |
| MMLU | 85.3% | #54 of 131 | Lab model card | |
| BOUQuET | 28.5 | #8 of 18 | Official leaderboard | |
| SuperGPQA | 44.9% | #30 of 59 | Official leaderboard | |
| LMArena (Elo) | 1,318 | #213 of 335 | Official leaderboard | |
| SEA-HELM | 60.6 | #50 of 71 | Official leaderboard | |
| Wer wird Millionär? LLM Benchmark | 296.0 | #55 of 77 | Official leaderboard | |
| AraGen | 30.6% | #30 of 42 | Official leaderboard | |
| MMMLU | 75.7% | #53 of 71 | Lab model card | |
| Humanity's Last Exam | 17.3% | #117 of 154 | Lab model card | |
| Last Translation Benchmark | 1.32 | #28 of 36 | Official leaderboard | |
| EQ-Bench 3 | 985 | #64 of 78 | Official leaderboard | |
| AgentIF (2026-10 setting) | 55% | #21 of 24 | Official leaderboard | |
| StudyArena Text | 1,183 | #75 of 85 | Official leaderboard | |
| SimpleQA | 6.70% | #55 of 60 | Lab model card | |
| IFEval | 73.2% | #56 of 60 | Official leaderboard | |
| SimpleQA Verified | 5.95% | #53 of 56 | Official leaderboard | |
| Judgemark v2.1 | 42.8 | #40 of 42 | Official leaderboard | |
| MultiNRC | 10.4% | #37 of 38 | Official leaderboard | |
| FACTS Parametric | 2.90% | #44 of 45 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 21.2 | #118 of 119 | Official leaderboard | |
| Creative Writing v3 | 40.5 | #119 of 120 | Official leaderboard | |
| FACTS Search | 34.8% | #31 of 31 | Official leaderboard | |
| Structured Output Benchmark | 73.2% | #19 of 19 | Official leaderboard | |
| IndicGenBench | 34.5 | #13 of 13 | Official leaderboard | |
| SimpleQA Hallucination Rate | 91.4% | #12 of 12 | Lab model card | |
| PersonQA | 15.5% | #9 of 9 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIME 2024 | 96% | #2 of 88 | Lab model card | |
| AIME 2025 | 98.7% | #13 of 143 | Lab model card | |
| MATH-500 | 94.2% | #13 of 38 | Official leaderboard | |
| ChemBench | 61.2% | #10 of 24 | Official leaderboard | |
| GPQA Main | 59.4% | #27 of 60 | Official leaderboard | |
| HMMT 2025 | 76.7% | #27 of 53 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 65.3% | #120 of 212 | Epoch AI run | |
| GPQA Diamond | 60.8% | #136 of 229 | Epoch AI run | |
| DTBench | 68% | #118 of 180 | Official leaderboard | |
| SMT 2025 | 83% | #27 of 40 | Official leaderboard | |
| CMIMC 2025 | 73.8% | #23 of 33 | Official leaderboard | |
| BRUMO 2025 | 86.7% | #30 of 42 | Official leaderboard | |
| SciArena | 928 | #28 of 34 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Codeforces rating | 2,516 | #11 of 47 | Lab model card | |
| Multi-LCB | 59.8% | #3 of 11 | Official leaderboard | |
| AIRS-Bench | 0.40 | #2 of 5 | Official leaderboard | |
| WeirdML | 40.9% | #75 of 112 | Official leaderboard | |
| SWE-bench Verified | 60.7% | #108 of 155 | Lab model card | |
| OpenAI PRs | 37% | #10 of 12 | Lab model card | |
| Aider Polyglot | 34.2% | #27 of 30 | Lab model card | |
| ALE-Bench | 942 | #115 of 126 | Official leaderboard | |
| DesignArena Website | 845 | #145 of 151 | Official leaderboard | |
| PaperBench | 7.00% | #15 of 15 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MCP-Bench | 0.65 | #6 of 15 | Official leaderboard | |
| LisanBench | 569.0 | #65 of 133 | Official leaderboard | |
| AppWorld Test-Normal | 76.2% | #7 of 12 | Official leaderboard | |
| BalatroBench | 5.80 | #11 of 16 | Official leaderboard | |
| lmgame-Bench | 17.4 | #17 of 23 | Official leaderboard | |
| LLM Chess | -232 | #98 of 123 | Official leaderboard | |
| tau-bench | 54.8% | #19 of 23 | Lab model card | |
| SnorkelWordle | 65.9% | #16 of 19 | Official leaderboard | |
| PinchBench | 48.4% | #43 of 48 | Official leaderboard | |
| PinchBench 2.0 | 36.3% | #54 of 57 | Official leaderboard | |
| Terminal-Bench 2.0 | 3.37% | #43 of 43 | Official leaderboard | |
| Terminal-Bench | 3.37% | #18 of 18 | tbench.ai, via Epoch AI | |
| AppWorld Test-Challenge | 15.3% | #12 of 12 | Official leaderboard | |
| tau-bench Retail | 54.8% | #2 of 2 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CorpFin v2 | 53.1% | #91 of 119 | Official leaderboard | |
| ProfBench | 42.3% | #56 of 72 | Official leaderboard | |
| LEXam | 32.1% | #24 of 30 | Official leaderboard | |
| LegalBench | 70.8% | #113 of 138 | Official leaderboard | |
| TaxEval v2 | 63.7% | #108 of 129 | Official leaderboard | |
| BRIDGE | 29.1 | #35 of 39 | Official leaderboard | |
| FailureSensorIQ | 43.5% | #22 of 24 | Official leaderboard | |
| CaseLaw v2 | 43.8% | #53 of 53 | Official leaderboard | |
| HealthBench | 42.5% | #31 of 31 | Lab model card | |
| HealthBench Hard | 10.8% | #21 of 21 | Lab model card | |
| HealthBench Consensus | 82.6% | #17 of 17 | Lab model card | |
| StockBench | 14.0 | #13 of 13 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CaP-X (CaP-Bench) | 7.50% | #9 of 10 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIR-Bench 2024 | 86% | #10 of 82 | Official leaderboard | |
| Anthropic Red Team | 99.7% | #12 of 75 | Official leaderboard | |
| MolecularIQ | 35.8% | #3 of 17 | Official leaderboard | |
| HELM Safety | 96.4 | #16 of 75 | Official leaderboard | |
| SimpleSafetyTests | 100% | #19 of 75 | Official leaderboard | |
| FORTRESS | 17.6% | #25 of 73 | Official leaderboard | |
| Tacit Knowledge and Troubleshooting | 73% | #12 of 30 | Lab model card | |
| Lakera B3 (Backbone Breaker Benchmark) | 55% | #18 of 35 | Official leaderboard | |
| BioTIER Refuse | 39.3% | #38 of 69 | Official leaderboard | |
| Capture the Flag (Collegiate) | 17% | #8 of 13 | Lab model card | |
| Multimodal Troubleshooting Virology | 50% | #19 of 27 | Lab model card | |
| BioTIER Permit | 97.9% | #55 of 69 | Official leaderboard | |
| BrowseSafe++ | 71% | #12 of 15 | Official leaderboard | |
| Human Pathogen Capabilities Test v2 (HPCT v2) | 37.9% | #57 of 71 | Official leaderboard | |
| nocot-bench (NCRI) | 54.8 | #157 of 185 | Official leaderboard | |
| KORA | 16.1% | #38 of 42 | Official leaderboard | |
| Capture the Flag (Professional) | 2.00% | #20 of 21 | Lab model card | |
| ProtocolQA Open-Ended | 16% | #27 of 27 | Lab model card | |
| TroubleshootingBench | 19% | #20 of 20 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HELM Capabilities | 67.4 | #27 of 60 | Official leaderboard | |
| ACCoRD | 61.9 | #84 of 152 | Official leaderboard | |
| Conceptual Reasoning Index | 30.4 | #116 of 151 | Official leaderboard | |
| LMCA | 14.5% | #126 of 147 | Official leaderboard |
The Self-Reported Index fits one score to the 16 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 130.0 to 136.8. How we compute it.
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.
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)
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 22 of them aren't in the launch post.