OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released two new reasoning models, o3 and o4-mini, to the public on April 16, 2025, making them available through ChatGPT and the company's API. The release marks the first time OpenAI has offered its o3 model, previously available only in limited research previews, to general users and developers.

The Models

O3 is OpenAI's most capable reasoning model to date, according to the company. It is designed to handle complex, multi-step tasks in mathematics, science, coding, and logical analysis. O4-mini is a smaller, faster, and lower-cost variant intended for developers who require high reasoning performance at greater throughput and reduced operating expense.

Both models support tool use, including web search, code execution via the Python interpreter, and image analysis. OpenAI stated that o4-mini achieves performance competitive with larger models on several benchmarks while operating at significantly lower inference cost.

On the AIME 2024 mathematics benchmark, a standard test for advanced reasoning systems, o3 scored 96.7 percent. On SWE-bench Verified, a software engineering benchmark measuring the ability to resolve real-world GitHub issues, o3 achieved 69.1 percent. O4-mini scored 93.4 percent on AIME 2024.

Access and Pricing

O3 is available to ChatGPT Plus, Pro, and Team subscribers. O4-mini is available to Plus, Pro, Team, and Free tier users, with Free tier access subject to usage limits. Both models are accessible through the OpenAI API.

For API customers, o4-mini is priced at $1.10 per million input tokens and $4.40 per million output tokens. O3 is priced at $10 per million input tokens and $40 per million output tokens. OpenAI also introduced a feature called extended thinking with tool use, which allows both models to use tools such as web search and code execution during their internal reasoning process before producing a final answer.

Background

OpenAI introduced its o-series reasoning models in September 2024 with the release of o1, followed by o1-mini. The series is distinct from the company's GPT line in that the models are trained to spend additional compute time reasoning through problems before generating responses, a technique OpenAI refers to as chain-of-thought reasoning.

O3 was first announced by OpenAI in December 2024 and demonstrated strong performance on several academic and professional benchmarks at that time. The model attracted attention for its score on the ARC-AGI benchmark, a test designed to measure general fluid intelligence in AI systems, where it reached 87.5 percent in a high-compute configuration.

The public release follows a period during which access to o3 was limited to select researchers and safety evaluators. OpenAI stated that the model underwent extensive safety testing prior to general availability, including evaluations for potential misuse in the development of biological, chemical, nuclear, and radiological weapons.

What It Means in Practice

Developers building applications that require multi-step problem solving, scientific analysis, or automated software engineering can now access both models through the standard API without joining a waitlist. The inclusion of tool use during the reasoning process, rather than only after a final answer is produced, represents a change in how the models interact with external data sources during inference.

Enterprises on OpenAI's business and enterprise ChatGPT plans will also receive access, with the company indicating that o3 will be added to those tiers in the coming weeks.

OpenAI has said it plans to release additional models in the o-series later in 2025, with further details expected at the company's upcoming developer events.