OpenAI · model
GPT-OSS 20B benchmark scores
As of 2026-11-21, GPT-OSS 20B (OpenAI) has measured scores on 19 benchmarks and estimated scores on 80 more.
| Benchmark | Score | Source |
|---|---|---|
| AA-GPQA Diamond | 68.8% | measured |
| AA-HLE | 11.0% | measured |
| AA-SciCode | 38.9% | measured |
| AIME26 | 90.9% | estimated ± 4.3 pp, medium confidence |
| Artificial Analysis Intelligence Index | 9.8% | estimated ± 3.2 pp, medium confidence |
| CritPt | 1.4% | measured |
| GDPval-AA | 0.0% | measured |
| GPQA Diamond | 71.7% | estimated ± 1.2 pp, high confidence |
| HLE | 10.9% | measured |
| IFBench | 63.7% | estimated ± 6.6 pp, medium confidence |
| AA-LCR | 34.7% | measured |
| LiveCodeBench v6 | 71.9% | estimated ± 4.7 pp, high confidence |
| AA-Omniscience Accuracy | 16.5% | estimated ± 2.4 pp, high confidence |
| SciCode | 39.1% | estimated ± 3.1 pp, high confidence |
| BrowseComp | 8.5% | estimated ± 4.2 pp, medium confidence |
| HLE w/ tools | 26.8% | estimated ± 4.1 pp, low confidence |
| IFEval | 92.5% | estimated ± 2.8 pp, high confidence |
| LongBench v2 | 51.3% | estimated ± 5.5 pp, medium confidence |
| τ²-bench results | 60.2% | measured |
| AA Coding Index | 20.7% | measured |
| APEX-Agents | 0.0% | estimated ± 9.4 pp, low confidence |
| APEX-Agents-AA | 0.7% | measured |
| BioMysteryBench (human-difficult) | 20.9% | estimated ± 6.8 pp, low confidence |
| DeepSearchQA | 26.7% | estimated ± 4.5 pp, medium confidence |
| SWE-bench Verified | 60.7% | measured |
| BFCL v4 | 53.9% | estimated ± 2.7 pp, high confidence |
| CyberGym | 46.6% | estimated ± 5.3 pp, low confidence |
| JobBench | 0.0% | estimated ± 8.2 pp, low confidence |
| SWE-bench Pro | 37.1% | estimated ± 4.5 pp, high confidence |
| WideResearch | 0.0% | estimated ± 7.8 pp, low confidence |
| AA Agentic Index | 1.4% | measured |
| MMLU-Pro | 72.3% | estimated ± 4.0 pp, high confidence |
| AA-IFBench | 65.1% | measured |
| AIME 2024 | 92.1% | measured |
| BBH | 90.2% | estimated ± 2.4 pp, low confidence |
| FailureSensorIQ | 43.5% | measured |
| GPQA | 71.5% | measured |
| MMLU | 85.3% | measured |
| Vals GPQA Diamond | 67.9% | estimated ± 3.2 pp, high confidence |
| Vals LiveCodeBench | 75.8% | estimated ± 5.7 pp, low confidence |
| Vals MMLU-Pro | 79.6% | estimated ± 2.3 pp, high confidence |
| MATH-500 | 97.9% | estimated ± 1.2 pp, medium confidence |
| MMMLU | 86.8% | estimated ± 0.7 pp, medium confidence |
| AIME 2025 | 91.7% | measured |
| AA MMLU-Pro | 85.3% | estimated ± 1.2 pp, low confidence |
| Gert Labs | 33.6% | estimated ± 8.3 pp, medium confidence |
| AA Harvey LAB v1.0 | 79.4% | estimated ± 1.2 pp, low confidence |
| AA Tau3 Banking | 16.8% | estimated ± 3.7 pp, medium confidence |
| ARC-AGI-1 | 24.7% | estimated ± 6.7 pp, low confidence |
| ARC-AGI-2 | 0.0% | estimated ± 10.9 pp, low confidence |
| CursorBench 3.2 | 36.0% | estimated ± 5.3 pp, low confidence |
| FrontierCode 1.1 Main | 0.0% | estimated ± 6.6 pp, low confidence |
| FrontierSWE v2 | 4.6% | estimated ± 14.2 pp, low confidence |
| MazeBench | 0.4% | estimated ± 5.5 pp, low confidence |
| OSWorld-Verified | 55.4% | estimated ± 8.4 pp, low confidence |
| Vals SWE-bench | 44.6% | estimated ± 5.5 pp, medium confidence |
| CWE-Bench-AA | 0.0% | estimated ± 10.1 pp, low confidence |
| AA Cyber Index | 1.0% | estimated ± 9.2 pp, low confidence |
| DeepsecBench-AA | 0.0% | estimated ± 8.9 pp, low confidence |
| AA ITBench | 33.3% | estimated ± 4.0 pp, medium confidence |
| Bug Hunt Bench | 13.9% | estimated ± 10.5 pp, low confidence |
| CursorBench 4.0 | 25.6% | estimated ± 4.9 pp, medium confidence |
| CWE-bench v1 | 0.0% | estimated ± 6.4 pp, low confidence |
| DeepSWE | 28.7% | estimated ± 7.5 pp, low confidence |
| DRACO | 0.0% | estimated ± 2.0 pp, low confidence |
| FrontierCode 2.0 | 26.1% | estimated ± 9.3 pp, low confidence |
| HLE w/o tools | 11.5% | estimated ± 5.3 pp, medium confidence |
| PostTrainBench v1.1 | 25.3% | estimated ± 7.4 pp, low confidence |
| Toolathlon-Verified | 5.5% | estimated ± 2.3 pp, low confidence |
| VulcanBench Frontier v4 | 68.5% | estimated ± 6.8 pp, low confidence |
| GDP.pdf | 17.0% | estimated ± 4.6 pp, medium confidence |
| ARC-AGI-3 | 1.3% | estimated ± 9.8 pp, low confidence |
| ARC-AGI-3 Provider Adapter | 2.7% | estimated ± 7.3 pp, low confidence |
| Vibe Code Bench | 27.7% | estimated ± 6.5 pp, low confidence |
| Hermes Bench | 0.0% | estimated ± 2.9 pp, medium confidence |
| Hermes Index | 0.0% | estimated ± 8.7 pp, low confidence |
| C-Eval | 92.2% | estimated ± 1.1 pp, medium confidence |
| HMMT Feb 2025 | 94.9% | estimated ± 2.9 pp, high confidence |
| HMMT Feb 2026 | 76.2% | estimated ± 11.0 pp, low confidence |
| MCP Atlas | 64.6% | estimated ± 7.8 pp, low confidence |
| MMLU-Redux | 90.9% | estimated ± 1.7 pp, high confidence |
| NL2Repo | 12.1% | estimated ± 4.7 pp, medium confidence |
| QwenClawBench | 55.9% | estimated ± 6.0 pp, low confidence |
| SuperGPQA | 52.5% | estimated ± 7.2 pp, medium confidence |
| Toolathlon | 27.0% | estimated ± 6.3 pp, low confidence |
| AA LiveCodeBench | 84.0% | estimated ± 4.7 pp, low confidence |
| HealthBench Hard | 12.3% | estimated ± 8.6 pp, low confidence |
| MMLU-Pro (Arcee) | 82.4% | estimated ± 4.0 pp, medium confidence |
| React Native Evals | 71.0% | measured |
| ResearchClawBench | 2.8% | estimated ± 1.9 pp, medium confidence |
| SWE-Rebench | 57.0% | estimated ± 3.6 pp, low confidence |
| BioMysteryBench (human-solvable) | 81.7% | estimated ± 0.6 pp, low confidence |
| ScarfBench | 24.4% | estimated ± 9.1 pp, low confidence |
| SimpleQA | 22.9% | estimated ± 7.8 pp, medium confidence |
| LiveCodeBench | 66.5% | estimated ± 14.6 pp, low confidence |
| OpenHarmony Bench v1.0 | 51.0% | estimated ± 2.4 pp, medium confidence |
| MRCRv2 | 22.8% | estimated ± 11.3 pp, low confidence |
| MMLongBench-Doc | 57.3% | estimated ± 2.6 pp, low confidence |
| BioBasicsBench | 33.9% | estimated ± 10.7 pp, low confidence |