OpenAI Model Spec Evals: leaderboard
OpenAI's 596 prompts on the chain of command, refusals, clarifying questions, sensitive topics, honesty and tone, graded by a model against the OpenAI Model Spec; the score is the overall compliance rate. OpenAI evaluated only its own models, unlike the independent constitution audit.
Metric: Overall Model Spec compliance rate (%). Source: alignment.openai.com. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (Thinking) | 89 | #167 (GPT-5) |
| 2 | GPT-5.4 (Thinking) | 87 | #123 (GPT-5.4) |
| 3 | GPT-5.3 Instant | 84 | #130 |
| 4 | O3 | 80 | #228 |
| 5 | GPT-4o | 72 | #741 |
Interactive version: theaggregate.ai/benchmark?slug=openai-model-spec-evals · How It Works · Data refreshed daily, snapshot 2026-11-13.