o1-preview and o1-mini
First model trained with large-scale reinforcement learning to think in a long hidden chain of thought before answering.
Spends extra inference compute on reasoning tokens that are billed but hidden (for safety and competitive reasons). Large jumps over GPT-4o on AIME, GPQA and Codeforces; text-only with no system prompt or streaming at launch. o1-mini is a cheaper, coding-focused variant.
- Date
- Thursday, 12 September 2024
- Lab
- OpenAI
- Kind
- model
- Access
- closed API
- Price
- o1-mini about 80% cheaper than o1-preview (Wikipedia)
Figures
| Measure | Value | Measured by |
|---|---|---|
| AIME 2024 | 83% for o1 vs 13% for GPT-4o launch figure for o1 (multi-sample setting); o1-preview scored lower; per Wikipedia summary of OpenAI post | company |
| Codeforces | 89th percentile competitive programming; company-reported | company |
API access initially limited to Tier 5 developers (Simon Willison). The multi-sample qualifier on the 83% AIME figure is recalled from the launch post and should be re-checked against the primary page.
Sources
- simonwillison.net/2024/Sep/12/openai-o1/
- en.wikipedia.org/wiki/OpenAI_o1
- developers.openai.com/api/docs/changelog
This record was checked against its sources on 6 October 2026. How we check