GPT-OSS-20B (Medium): benchmark results
GPT-OSS-20B evaluated at the medium reasoning-effort setting. Provider: OpenAI. Released 2025-08-05. Access: Open.
Aggregate Rating 1547 ± 8, rank #888 of 2168 rated models, from 412 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| PhantomBench - Terms (Existence) | 2.42 | Hallucination rate (%): share of answers that do not abstain | 100 |
| EuroEval Polish Summarization - PSC | 25.35 | chrF3++ (%) | 95.9 |
| EuroEval Slovak NLU - UNER SK | 65.77 | Micro F1 excluding MISC (%) | 95.8 |
| Multi-LCB - Pass@5 T0.6 - Go | 79.5 | Pass@5 (%) | 95.8 |
| Multi-LCB - Pass@5 T0.6 - Ruby | 79.1 | Pass@5 (%) | 95.8 |
| Multi-LCB - Pass@5 T0.6 - Rust | 79.8 | Pass@5 (%) | 95.8 |
| Multi-LCB - Pass@5 T0.6 - TypeScript | 82.5 | Pass@5 (%) | 95.8 |
| Swallow - English MT-Bench - Extraction | 83.4 | Judge Score (normalized, %) | 95.5 |
| EuroEval Latvian NLU - Latvian Twitter Sentiment | 51.88 | Matthews correlation coefficient (%) | 95.3 |
| PhantomBench - Entities (Existence) | 3.93 | Hallucination rate (%): share of answers that do not abstain | 93.3 |
| EuroEval Spanish NLU - ScaLA ES | 41.52 | Matthews correlation coefficient (%) | 92 |
| Multi-LCB - Pass@5 T0.6 - C++ | 81.5 | Pass@5 (%) | 91.7 |
Interactive version: theaggregate.ai/model?slug=gpt-oss-20b-medium · How It Works · Data refreshed daily, snapshot 2026-10-20.