OpenAI drops o1-mini: 80% cheaper reasoning that nearly matches o1

OpenAI just rolled out a model that might actually change the cost calculus for developers building reasoning-heavy applications. o1-mini is a smaller, distilled version of the company’s reasoning models, and it’s arriving at 80% less than o1-preview for API users. That’s not a typo. The model is available starting today for tier 5 API accounts, along with ChatGPT Plus, Team, Enterprise, and Edu subscribers, who will also see higher rate limits and lower latency.

The headline number tells a compelling story on its own, but the benchmarks fill in the picture. On the AIME high school math competition, o1-mini scored 70.0% — competitive with the full o1 at 74.4% and miles ahead of o1-preview’s 44.6%. That score lands it among the top 500 high school math students in the US. Coding tells a similar story: a 1650 Elo on Codeforces, putting it at the 86th percentile of competitive programmers and just a hair below o1’s 1673. OpenAI trained it using the same high-compute reinforcement learning pipeline as the larger models, but optimized it specifically for STEM reasoning during pretraining.

The trade-off, predictably, is breadth. o1-mini stumbles on tasks requiring broad factual knowledge. On MMLU it lags behind GPT-4o, and on GPQA science questions it falls short of o1-preview. OpenAI is candid that its non-STEM knowledge — dates, biographies, trivia — is comparable to small models like GPT-4o mini. That’s the bargain you make for speed and cost efficiency.

OpenAI says safety wasn’t shortchanged. The model went through the same alignment training, external red-teaming, and preparedness evaluations as o1-preview, scoring 59% higher on jailbreak robustness than GPT-4o. In human preference tests, raters favored o1-mini over GPT-4o in reasoning-heavy domains but not in language-focused tasks. One concrete example showed both o1-mini and o1-preview solving a word reasoning question that stumped GPT-4o, with o1-mini reaching the answer 3-5x faster. For developers who need cheap reasoning and don’t need an encyclopedia, this might be the sweet spot.