OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released two new reasoning models, o3 and o4-mini, making them publicly available through ChatGPT and its API. The release marks the first time OpenAI has offered its o3 model, previously available only to researchers and select testers, to the general public and commercial developers.

Background

OpenAI has maintained two parallel product lines since late 2023: its GPT series, designed for fast general-purpose language tasks, and its o-series, designed to spend additional compute time on multi-step reasoning before producing a response. The o-series models are intended for tasks that benefit from deliberate, structured thinking, including mathematics, coding, and scientific analysis.

The original o1 model launched in September 2024, followed by o1-mini and o3-mini in subsequent months. o3 was first described publicly in December 2024 and drew significant attention after benchmark results showed strong performance on tests including the ARC-AGI evaluation, which measures abstract reasoning. At the time of that announcement, OpenAI did not set a firm public release date.

o4-mini is a new entry in the series. OpenAI has described it as a smaller, faster, and more cost-efficient model in the o-series line, positioned for developers who require reasoning capabilities at lower latency and cost than the full o3 model.

What the Models Do

Both o3 and o4-mini use a technique OpenAI calls "extended thinking," in which the model generates an internal chain of reasoning steps before delivering a final answer. This approach differs from standard autoregressive generation in that the model allocates variable amounts of compute depending on problem complexity.

OpenAI has stated that o3 achieves top scores on several established benchmarks, including AIME 2024, a mathematics competition test, and Codeforces, a competitive programming evaluation. The company has also reported strong results on SWE-bench, which measures performance on real-world software engineering tasks drawn from GitHub repositories.

o4-mini is reported to outperform the previously released o3-mini on several of the same benchmarks while operating at reduced cost. OpenAI has indicated that o4-mini will be available at a lower price tier through its API than o3.

Both models also support tool use, including web search and the ability to run and interpret code, capabilities OpenAI refers to collectively as "agentic" functions. This allows the models to retrieve information and execute steps autonomously within a defined task.

Access and Pricing

ChatGPT Plus, Pro, and Team subscribers can access o3 and o4-mini through the ChatGPT interface. API access is available to developers through OpenAI's platform. Specific per-token pricing details were published in OpenAI's API documentation at the time of release.

OpenAI has also indicated that o3 is available to free-tier ChatGPT users on a limited basis, subject to usage caps.

Market Context

The release comes during a period of active competition in the reasoning model segment. Google DeepMind has released Gemini 2.5 Pro, which the company has described as its most capable reasoning model. Anthropic offers Claude 3.7 Sonnet with an extended thinking mode. Meta has released Llama 4, which includes reasoning-oriented capabilities in its open-weight format.

OpenAI holds a leading position in API-based AI services by reported revenue and developer adoption, though precise market share figures are not publicly available.

OpenAI has not announced a timeline for subsequent models in the o-series, though the company has previously indicated that an o3-pro variant and additional agentic products are in development.