FrontierMath Tiers 1-3 - OpenAI internal run
Math · 2025-08-07
FrontierMath Tier 1-3 accuracy as run by OpenAI itself for its 2025 release posts (Python tool), on OpenAI's copy of the problem set rather than Epoch AI's private-split harness: OpenAI's GPT-5 (high) figure of 26.3% differs from Epoch's 32.4% run of the same model. Kept separate from the Epoch v1/v2 private rows; diagnostic. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5 Pro | 32.1 |
| GPT-5.1 | 26.7 |
| GPT-5 | 26.3 |
| GPT-5 Mini | 22.1 |
| o3 | 15.8 |
| GPT-5 Nano | 9.6 |