OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released its o3 and o4-mini reasoning models to the public in April 2025, making the company's most advanced reasoning systems available beyond its research and enterprise preview programs. The release marks the broadest public availability of OpenAI's third-generation reasoning model family, which the company has described as its most capable to date for complex, multi-step problem solving.

The models are accessible through OpenAI's API and through ChatGPT for subscribers at relevant tier levels. o4-mini is positioned as a smaller, faster, and lower-cost variant designed to deliver strong reasoning performance at reduced computational expense compared to the full o3 model.

Background

OpenAI introduced its "o" series reasoning models in late 2024, beginning with o1, which was designed to spend additional compute time working through problems before producing a response. This approach, sometimes described as chain-of-thought reasoning at inference time, was developed to improve performance on tasks requiring logic, mathematics, coding, and scientific analysis.

o3 was first announced by OpenAI in December 2024 and drew significant attention after the company reported benchmark results showing substantial performance gains over o1 on evaluations including the ARC-AGI benchmark, a test designed to measure general reasoning ability that prior AI systems had struggled to perform well on. OpenAI reported o3 achieved a score of approximately 87.5 percent on the ARC-AGI semi-private evaluation set under high-compute conditions.

The o4-mini model follows a pattern OpenAI established with o1-mini, offering a distilled variant intended to lower costs for developers and businesses that require reasoning capabilities but face budget or latency constraints.

What the Models Do

o3 is a full-scale reasoning model intended for tasks where accuracy and depth of analysis take priority over response speed. OpenAI has stated it performs at or near the top of publicly reported benchmarks across mathematics, competitive programming, and scientific question answering.

o4-mini is optimized for speed and cost efficiency while retaining strong reasoning performance relative to its size. OpenAI has indicated o4-mini performs competitively against larger models on many reasoning tasks, making it suitable for high-volume API applications.

Both models support tool use, including web search and code execution in applicable environments, extending their ability to work through problems that require retrieving or processing external information.

Access and Pricing

Developers can access both models through the OpenAI API under the company's standard usage-based pricing structure. o3 is priced at a premium tier reflecting its computational requirements. o4-mini carries lower per-token costs, consistent with OpenAI's tiered model strategy.

ChatGPT Plus, Pro, and Team subscribers have access to both models through the ChatGPT interface. Enterprise and Education account access follows separately negotiated terms.

Market Context

The public release of o3 and o4-mini comes as competition in the reasoning model segment has intensified. Google DeepMind's Gemini 2.5 Pro, released in early April 2025, has also reported strong benchmark results on reasoning evaluations. Anthropic's Claude model family includes extended thinking features positioned for similar use cases.

OpenAI, which was founded in 2015 and restructured as a capped-profit entity, has reported that its annualized revenue reached approximately 3.4 billion dollars as of early 2025. The company counts more than 300 million weekly active ChatGPT users according to figures it has made public.

The reasoning model segment is viewed within the industry as a distinct product category from standard language model completions, with separate demand from enterprises in legal, financial, scientific, and software development sectors.

OpenAI has indicated it will continue releasing updates to its model lineup throughout 2025, with additional models in the o-series and GPT-4 family expected later in the year.