OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released its o3 and o4-mini reasoning models to the general public, making the company's latest generation of deliberative AI systems available beyond closed research and developer previews. The release marks the broad public availability of models OpenAI describes as its most capable reasoning systems to date, designed to work through complex, multi-step problems across mathematics, science, coding, and logical analysis.

Background

OpenAI has maintained a tiered release strategy for its reasoning model line since introducing the o1 series in late 2024. That line, distinct from the company's GPT series, is built around a technique in which the model spends additional compute time generating internal reasoning steps before producing a final answer. This approach, sometimes called chain-of-thought or deliberative reasoning, is intended to improve accuracy on tasks that require sequential logic rather than pattern matching alone.

The o3 model was first described publicly by OpenAI in December 2024, when the company shared benchmark results but did not release the model for general use. At that time, o3 posted top scores on several established AI benchmarks, including the ARC-AGI test, which measures performance on novel reasoning tasks not easily solved by memorisation. The o4-mini is a smaller, more computationally efficient variant intended to deliver strong reasoning performance at lower cost and latency.

Prior to this public release, access to o3 and o4-mini had been limited to select researchers, safety evaluators, and API developers through a staged rollout process OpenAI uses to assess model behaviour before wider deployment.

Key Capabilities

According to OpenAI, o3 and o4-mini both support tool use, meaning the models can call external functions, browse the web, execute code, and process images as part of their reasoning chains. This integration of tool use within the reasoning process is a technical distinction from earlier reasoning models, which operated primarily on text input without native tool-calling during the chain-of-thought phase.

OpenAI has stated that o4-mini, despite its smaller parameter footprint, performs competitively with larger models on mathematical and coding benchmarks, making it suited for high-volume applications where inference cost is a factor. Both models are available through the OpenAI API and within the ChatGPT interface for subscribers on Plus, Pro, and Team plans.

Market Context

The release comes as competition in the reasoning model segment has intensified. Google DeepMind has released its Gemini 2.5 Pro model with reasoning capabilities, and several open-source and commercially available models from other developers have incorporated similar deliberative approaches. The timing also follows OpenAI's release of the GPT-4.1 API, which the company positioned as a coding and instruction-following model rather than a reasoning-focused system, indicating OpenAI is maintaining separate product lines for different use cases.

OpenAI, backed by Microsoft and valued at approximately 300 billion dollars following its most recent funding discussions, has described its reasoning model line as central to its longer-term research direction toward systems capable of autonomous, agentic task completion.

What It Means in Practice

Developers with API access can now integrate o3 and o4-mini into applications requiring multi-step problem solving, scientific computation, or code generation and debugging. Pricing for the models through the API is tiered by token volume, with o4-mini carrying a lower per-token cost than o3. Enterprise customers on custom contracts may have access to additional rate limits and deployment options.

ChatGPT users on paid plans can access both models directly through the model selector within the standard chat interface, with usage subject to rate limits that vary by subscription tier.

OpenAI has indicated it will continue publishing safety and evaluation documentation for o3 and o4-mini as part of its standard model release process, with additional technical detail expected in the weeks following the public launch.