GPT-5 (Medium): benchmark results
GPT-5 evaluated at the medium reasoning-effort setting. Provider: OpenAI. Released 2025-08-07. Access: API.
Aggregate Rating 1713 ± 7, rank #318 of 2168 rated models, from 315 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| FIRE-Bench - Cross-Domain | 30.2 | Claim-level F1 (%; atomic-claim entailment against the sourc | 100 |
| HAL Online Mind2Web | 42.33 | Accuracy (%) | 100 |
| HAL USACO | 69.06 | Accuracy (%) | 100 |
| IRB1K (Closed-Book) - Cross-Lingual | 31.5 | Correctness (%, 0-100): share of answers judged correct (mea | 100 |
| IRB1K (Closed-Book) - English | 41.9 | Correctness (%, 0-100): share of answers judged correct (mea | 100 |
| IRB1K (RAG) - Cross-Lingual | 77.6 | Correctness (%, 0-100): share of answers judged correct (mea | 100 |
| OdysseyArena - Energy Dispatch (Pass@4) | 40 | Success rate (%) with a task counted solved when any of four | 100 |
| OdysseyArena - Repo System (Pass@4) | 83.33 | Success rate (%) with a task counted solved when any of four | 100 |
| SecRepoBench (Standalone) - Pass@1 | 54.1 | Completions that pass the unit tests, pass@1 (%) | 100 |
| SecRepoBench (Standalone) - Secure-Pass@1 | 39.3 | Completions that pass the unit tests and the PoC security te | 100 |
| SimuScene (Pass@8) | 59.9 | Pass@8 accuracy (%): share of scenarios with at least one of | 100 |
| Step Game (Lechmazur) | 5.49 | TrueSkill μ | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-5-medium · How It Works · Data refreshed daily, snapshot 2026-10-20.