OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released two new reasoning-focused artificial intelligence models, o3 and o4-mini, making them available to the public through its API and consumer products. The release represents the company's latest generation of models designed specifically to handle complex, multi-step reasoning tasks across mathematics, science, coding, and general problem-solving.

Background

OpenAI has developed a separate line of reasoning-oriented models, distinct from its general-purpose GPT series, beginning with the o1 model released in late 2024. These models are trained using a technique that involves extended internal reasoning before producing a final response, a process the company refers to as "thinking." Earlier versions, including o1 and o3-mini, were made available on a limited or tiered basis before broader rollout.

The o3 model had been previously discussed by OpenAI in benchmark contexts, drawing significant attention after performance results on standardized tests circulated among researchers. The o4-mini designation follows the naming convention established with the earlier o3-mini, positioning it as a smaller, faster, and more cost-efficient variant within the same reasoning model family.

Model Capabilities and Access

OpenAI described o3 as the company's most capable reasoning model at the time of release, designed for tasks requiring sustained logical chains and high accuracy on technical subject matter. The o4-mini is positioned for use cases where lower latency and reduced cost are priorities while still delivering improved reasoning relative to prior compact models.

Both models are accessible through OpenAI's API, with o3 available to users on paid tiers and o4-mini offered at a lower price point suited to higher-volume applications. ChatGPT Plus and Pro subscribers also received access to both models through the consumer interface.

OpenAI stated that o4-mini performs competitively with larger models on several benchmarks while operating at a fraction of the computational cost, making it viable for developers building applications that require frequent inference calls.

Tool Use and Multimodal Input

A notable feature of both models is their integration with tool use during the reasoning process itself. Unlike prior iterations, o3 and o4-mini can invoke tools such as web search, code execution, and image analysis while constructing their internal reasoning chains, rather than only before or after generating a response. OpenAI indicated this allows the models to gather and incorporate new information mid-reasoning, which the company said improves accuracy on tasks requiring real-world data.

Both models also support image input as part of the reasoning process, extending multimodal capabilities into the reasoning model line for the first time in OpenAI's public releases.

Pricing and API Details

OpenAI published per-token pricing for both models at launch. The o3 model is priced at ten dollars per million input tokens and forty dollars per million output tokens. The o4-mini is priced at one dollar and ten cents per million input tokens and four dollars and forty cents per million output tokens. These figures place o4-mini among the more cost-accessible advanced reasoning options available through the OpenAI API at the time of release.

Competitive Context

The release comes during a period of rapid activity across the AI sector. Google DeepMind, Anthropic, and Meta have all released or updated major models in recent weeks. OpenAI also recently released GPT-4.1 through its API, targeting long-context applications with a one-million-token window. The o3 and o4-mini releases address a different segment, specifically tasks where step-by-step reasoning accuracy is prioritized over raw context length.

OpenAI is expected to continue expanding model access and updating pricing tiers as usage data from the initial rollout is collected.