OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released its o3 and o4-mini reasoning models to the public, making the company's most advanced reasoning systems broadly available for the first time. The release marks the first time OpenAI has offered models from its "o-series" reasoning line outside of limited research and enterprise previews.

Background

OpenAI's o-series models are designed around a process the company calls "chain-of-thought" reasoning, in which the model spends additional compute time working through problems in structured steps before producing a final response. The approach differs from the company's GPT-series models, which generate responses in a single forward pass.

The original o1 model was released in preview form in September 2024, followed by o1-mini, a smaller and faster variant. o3, the next generation in the series, was announced publicly in December 2024 but had been accessible only to safety researchers and select partners ahead of this broader release. o4-mini is a compact variant intended to deliver reasoning capabilities at lower cost and latency.

OpenAI has stated that o3 achieved high scores on several established benchmarks, including the ARC-AGI benchmark developed by researcher Francois Chollet, which measures a model's ability to solve novel visual reasoning tasks. The company reported that o3 scored approximately 87.5 percent on the ARC-AGI semi-private evaluation set under high-compute conditions, a figure that drew significant attention in the research community when disclosed earlier this year.

What the Models Do

o3 is positioned as OpenAI's most capable publicly available reasoning model and is intended for complex tasks including advanced mathematics, scientific reasoning, multi-step coding problems, and structured analysis. o4-mini is designed for use cases that require faster responses or lower operating costs, while still applying chain-of-thought processing.

Both models are accessible through OpenAI's API. o3 is available to users on paid tiers of ChatGPT, while o4-mini is available more broadly, including on ChatGPT's free tier with usage limits. Pricing details for API access were published on OpenAI's developer platform at the time of release.

One notable addition in the o4-mini and o3 release is support for tool use, including web browsing, code execution, and file analysis within the reasoning loop. Earlier o-series models had limited tool integration compared to GPT-series counterparts.

Market Context

The release comes as competition in the reasoning model segment has intensified. Google DeepMind released Gemini 2.5 Pro with enhanced reasoning capabilities in the same period, and Anthropic has promoted the reasoning performance of its Claude 3.7 Sonnet model. Chinese AI laboratory DeepSeek released its R1 reasoning model earlier in 2025, which drew widespread attention for matching or approaching the performance of larger Western models at significantly lower reported training costs.

OpenAI has positioned the o-series as distinct from its GPT-series, with the two lines targeting different use cases. The company has indicated that future model releases may integrate both generative and reasoning approaches more closely.

Access and Availability

Developers can access o3 and o4-mini through the OpenAI API using existing credentials. OpenAI has published a model card and system card for both releases, documenting evaluated risks, safety testing procedures, and known limitations. The company noted that o3 and o4-mini underwent its standard pre-deployment safety evaluations, including assessments for potential misuse in biological, chemical, radiological, and nuclear domains.

Rate limits and pricing tiers for both models are listed on OpenAI's platform documentation, with enterprise customers able to negotiate higher usage limits through direct agreements.

OpenAI has indicated that additional model releases are planned for later in 2025, including updates to its GPT-series and further iterations of the o-series reasoning line.