OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released its o3 and o4-mini reasoning models to paying subscribers on Wednesday, making the company's most capable reasoning-focused systems broadly available for the first time. The release follows months of limited research previews and positions the models as OpenAI's current flagship offerings for tasks requiring multi-step logical reasoning, mathematics, and coding.

Background

OpenAI introduced its "o" series of reasoning models in September 2024, beginning with o1, which was designed to spend additional computational time working through problems before producing a response. The series was framed as distinct from the company's GPT line, with the o models optimised for deliberate, step-by-step reasoning rather than fast generative output.

o3 was first announced in December 2024 and drew significant attention after OpenAI reported benchmark results showing the model achieved a score of 87.5 percent on the ARC-AGI benchmark, a test designed to measure general reasoning ability that prior AI systems had struggled to exceed. o3 also posted top results on several competitive mathematics and coding evaluations at the time of its preview announcement.

o4-mini is a smaller, more computationally efficient variant intended to deliver strong reasoning performance at lower cost and latency. OpenAI has previously used the "mini" designation for models that trade some capability for speed and reduced operational expense, making them more practical for high-volume or cost-sensitive applications.

What the Models Do

o3 and o4-mini both employ a chain-of-thought reasoning process in which the model generates internal reasoning steps before producing a final answer. This approach differs from standard large language model inference and is intended to improve accuracy on problems that benefit from structured problem decomposition.

A notable addition in both models is the ability to use tools during the reasoning process itself. According to OpenAI, the models can call tools such as web search and code execution mid-reasoning, rather than only at the start or end of a task. This allows the models to retrieve information or verify computations as part of their internal reasoning chain before delivering a response.

o4-mini with its "high" reasoning setting was reported by OpenAI to match or exceed o3 performance on several benchmarks, including competitive programming and science question sets, while consuming fewer computational resources.

Availability and Pricing

Both models are available through OpenAI's API and to ChatGPT Plus, Pro, and Team subscribers. ChatGPT Free users have access to o4-mini in a limited capacity. Enterprise and education tier access was also confirmed as part of the rollout.

API pricing for o3 was set at 10 dollars per million input tokens and 40 dollars per million output tokens. o4-mini is priced at 1.10 dollars per million input tokens and 4.40 dollars per million output tokens, continuing OpenAI's practice of offering a lower-cost reasoning option alongside its flagship model.

Competitive Context

The release comes as Google DeepMind, Anthropic, and other AI developers have also advanced their own reasoning-capable models. Google DeepMind released Gemini 2.5 Pro with enhanced reasoning capabilities in recent weeks. Anthropic's Claude 3.7 Sonnet introduced extended thinking features earlier in 2025. The reasoning model segment has become a primary competitive axis among frontier AI developers.

OpenAI also operates under a partnership with Microsoft, which distributes OpenAI models through its Azure OpenAI Service. Microsoft is expected to make o3 and o4-mini available through Azure following the public release.

What Happens Next

OpenAI has indicated that additional model releases are planned for 2025, with the company previously referencing GPT-5 and further updates to the o series on its public roadmap.