OpenAI ships o1 model with new developer tools

OpenAI moved o1 out of preview and into general availability in its API, two weeks after shipping the full model to ChatGPT alongside a $200-a-month subscription tier. The API release brought a set of developer tools that had not been available with o1-preview: function calling, so the model could call external tools and APIs as part of its reasoning; structured outputs, which constrained responses to a specified JSON schema; vision, letting the model reason over uploaded images; and a new reasoning_effort parameter that let developers trade cost and latency against how long the model deliberated before answering.

OpenAI reported that the version shipped, internally dated o1-2024-12-17, used on average 60% fewer reasoning tokens than o1-preview for a comparable request, and cited state-of-the-art scores on benchmarks including MATH and LiveCodeBench. Access at launch was limited to developers on OpenAI’s highest usage tier — those who had already spent over $1,000 on API calls — with broader rollout to follow over subsequent weeks; API access to the “o1 pro mode” available in ChatGPT Pro was not yet available and was described as coming soon.

The release completed the shift, begun with o1-preview in September, from a reasoning model shown mainly as a research and chat demonstration to one built into the tooling — function calling, schema-constrained output, an adjustable compute dial — that production applications required. It set the template OpenAI and competing labs followed with subsequent reasoning-model releases: a consumer-facing launch first, an API release with developer infrastructure to follow within weeks.