OpenAI releases o1, trading inference time for reasoning

OpenAI released o1-preview, a model trained with reinforcement learning to produce a long internal chain of thought before answering. The reported gains were concentrated in domains with verifiable answers: on a qualifying exam for the USA Mathematical Olympiad it solved 83% of problems against GPT-4o’s 13%, and it placed in roughly the 89th percentile on Codeforces competitive programming problems.

The significant claim was about scaling. OpenAI published curves showing accuracy rising smoothly with both the reinforcement-learning training compute and the amount of compute spent thinking at inference time. Since 2020 the field’s scaling story had been about pre-training; o1 described a second axis, independent of it, along which more compute could be converted into more capability. Coming amid persistent reporting that pre-training returns were flattening, this reframed the near-term outlook.

The model’s costs were real. It was slower and considerably more expensive, and OpenAI chose not to show users the raw chain of thought — citing the value of leaving it unconstrained by safety post-training, and, less prominently, competitive concerns. That decision made independent analysis of how the model reasoned largely impossible, and provoked warnings when users probing the hidden trace received policy violation notices.

The response was fast. Google shipped Gemini 2.0 Flash Thinking in December, Alibaba and DeepSeek published reasoning models, and within four months DeepSeek’s R1 reproduced much of the approach with open weights and a published method. The full o1 shipped in December 2024 and o3 was announced days later with a result on the ARC-AGI benchmark that had been considered years away.

“Test-time compute” became the organising idea of the following two years, and the principal argument that the trajectory had not slowed.