OpenAI internal API instruction following (hard) - reported
Chat & Writing · 2025-08-07
OpenAI's internal API instruction-following evaluation (hard split), in which models must follow difficult instructions derived from real developer feedback. The item set and grading are unpublished; diagnostic. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5 Mini | 65.8 |
| GPT-5 | 64.0 |
| GPT-5 Nano | 56.1 |
| GPT-4.5 | 54.0 |
| o1 (2024-12-17) | 51.3 |
| o3-mini | 50.0 |
| GPT-4.1 | 49.1 |
| o3 | 47.4 |
| GPT-4.1 Mini | 45.1 |
| O4 Mini | 44.7 |
| GPT-4.1 Nano | 31.6 |
| GPT-4o (2024-11-20) | 29.2 |
| GPT-4o Mini | 27.2 |