GPT-OSS-20B (High): benchmark results
GPT-OSS-20B evaluated at the high reasoning-effort setting. Provider: OpenAI. Released 2025-08-05. Access: Open.
Aggregate Rating 1514 ± 7, rank #1017 of 2168 rated models, from 534 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LLM Self-Modeling - Component-Attribution | 0.03 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - Flip-Decision | 0.08 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - Flip-Rate | -0.05 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - Self-Accuracy | 0.01 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| PersonalHomeBench - Information Grounding (Sole Reasoning) | 81.2 | Accuracy (%) on 1,000 information grounding questions (factu | 100 |
| MolecularIQ - Counting | 39.1 | Mean Accuracy (%) | 97.3 |
| MolecularIQ - Indexing | 32.7 | Mean Accuracy (%) | 94.6 |
| EuroEval Latvian NLU - Latvian Twitter Sentiment | 51.26 | Matthews correlation coefficient (%) | 93.5 |
| MolecularIQ - Overall | 35.7 | Mean Accuracy (%) | 91.9 |
| AA LiveCodeBench | 77.67 | Pass@1 (%) | 89.3 |
| MolecularIQ - Generation | 35.2 | Mean Accuracy (%) | 89.2 |
| EuroEval Spanish NLU - ScaLA ES | 39.3 | Matthews correlation coefficient (%) | 89.1 |
Interactive version: theaggregate.ai/model?slug=gpt-oss-20b-high · How It Works · Data refreshed daily, snapshot 2026-10-20.