Editor’s update, October 19, 2026: This article covers OpenAI’s January 31, 2025 launch of o3-mini. OpenAI now labels the model deprecated in the API and schedules o3-mini / o3-mini-2025-01-31 for shutdown on October 23, 2026. The January 2025 ChatGPT rollout, model-picker options, and message limits below are historical. See the official API deprecation schedule before planning a deployment.

Image: Official OpenAI o3-mini announcement page, originally published January 31, 2025, captured October 19, 2026. The page now includes an API deprecation notice; this is not a model or benchmark test.

OpenAI introduced o3-mini as a compact reasoning model aimed at science, mathematics, software development, and other technical problems. Previewed in December 2024 and released on January 31, 2025, it followed o1-mini with a focus on combining reasoning quality and lower latency. This page preserves that launch context rather than presenting o3-mini as OpenAI’s latest model.

What o3-mini introduced for developers

At launch, o3-mini supported function calling, Structured Outputs, developer messages, and streaming. These features helped developers request structured responses and connect model output to application workflows. They did not eliminate the need to validate generated code, tool arguments, or other outputs.

Developers could select three reasoning-effort settings:

  • Low: prioritize response speed when the task needed less reasoning.
  • Medium: balance reasoning quality and latency.
  • High: spend more effort on difficult multi-step problems, potentially increasing response time.

o3-mini did not support image inputs. Its search integration was described as an early ChatGPT prototype, so web access should not be treated as an inherent capability of every API call. OpenAI’s model reference separately documents API features, supported modalities, and limits.

The original ChatGPT and API rollout

The January 2025 release brought o3-mini to ChatGPT Plus, Team, and Pro users, with Enterprise access announced for February. It replaced o1-mini in the model picker. At that time, OpenAI increased the Plus and Team allowance from 50 o1-mini messages to 150 o3-mini messages per day.

Free users could try the reasoning model through the Reason option in the composer or by regenerating a response. Paid users also had an o3-mini-high option. The initial API rollout covered selected developers in the then-current usage tiers 3–5. These details explain the 2025 launch; they are not current account-entitlement instructions.

How to read the launch performance claims

The original announcement reported improvements over o1-mini on technical evaluations. Reasoning effort, tool access, evaluation scaffolding, and the tested dataset all affected the results. A benchmark result is not a guarantee that an application will achieve the same accuracy or speed.

  • Mathematics and science: OpenAI reported that medium effort was comparable to o1 on AIME, while high effort was comparable to o1 on GPQA Diamond.
  • Coding: Codeforces scores increased with reasoning effort. SWE-bench Verified results used 477 tasks and varied by the agent scaffold; the launch’s leading-performance language applies to models and tests available at that time.
  • FrontierMath: the reported result of over 32% solved on the first attempt, including more than 28% of T3 problems, used high reasoning effort and a Python tool. OpenAI explicitly called these numbers provisional. They should not be compared with the no-tool chart as if the conditions were identical.
  • Human evaluation: external testers preferred o3-mini over o1-mini 56% of the time and observed 39% fewer major errors on the difficult questions evaluated.

In OpenAI’s launch A/B testing, average response time was 7.7 seconds for o3-mini versus 10.16 seconds for o1-mini, about 24% faster. Those are test measurements rather than a latency promise for every prompt or deployment. The original announcement provides the comparison context.

Safety and remaining limitations

OpenAI described using deliberative alignment, preparedness evaluations, and external red-teaming. Its safety findings concerned specific evaluations; they did not establish that the model could never produce unsafe or incorrect answers. Technical outputs still require testing, and important factual answers need verification.

Current API lifecycle: plan beyond the launch model

As checked on October 19, 2026, OpenAI’s schedule lists October 23, 2026 as the API shutdown date for o3-mini and its January 31, 2025 snapshot. This is a scheduled future shutdown at the time of this update. OpenAI lists gpt-5.6-sol as the substitute in that notice. Review the current model catalog and test any replacement against the application’s own tasks before migrating.

o3-mini’s lasting significance is its role in the 2025 move toward smaller, configurable reasoning models. Its historical launch features and evaluations remain useful context, while present-day model selection depends on current documentation and lifecycle notices.