OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released two new reasoning models, o3 and o4-mini, making them available to ChatGPT Plus, Pro, and Team subscribers as well as API developers on April 16, 2025. The release marks the first time OpenAI has made its most advanced reasoning-class models broadly available outside of limited research and safety testing programs.

Background

OpenAI has developed a line of models it designates with the "o" prefix to distinguish them from its GPT series. These models are trained using reinforcement learning techniques designed to improve multi-step logical reasoning, mathematical problem-solving, and coding tasks. Earlier models in the series, o1 and o1-mini, were released in stages beginning in late 2024, initially restricted to select users before broader rollout.

The o3 model was first previewed publicly in December 2024, when OpenAI shared benchmark results showing performance gains on standardized tests including the ARC-AGI benchmark, a suite designed to measure general reasoning ability. At the time, OpenAI said o3 scored 87.5 percent on the ARC-AGI semi-private evaluation set, a result that drew significant attention from researchers monitoring progress in AI reasoning. The company did not release the model for public use at that time, citing the need for additional safety evaluation.

o4-mini is a smaller, faster, and lower-cost variant designed for deployment contexts where inference speed and API pricing are priorities. OpenAI has used this naming and sizing strategy previously, offering both full and mini versions of its o1 model to serve different use cases and cost profiles.

What the Models Do

Both o3 and o4-mini use a technique OpenAI describes as "thinking before responding," in which the model generates internal reasoning steps before producing a final answer. This approach differs from standard large language model inference, where output is generated token by token without a discrete reasoning phase. The visible reasoning trace can be inspected by users in some interface configurations.

OpenAI states that o3 shows improvements over prior models on tasks involving advanced mathematics, scientific reasoning, and multi-step coding problems. The company published benchmark comparisons at release showing o3 outperforming o1 on several internal and third-party evaluations. o4-mini is reported to exceed o1-mini on similar benchmarks while maintaining lower latency and reduced API cost per token.

Both models also support tool use, including web search and code execution through OpenAI's existing function-calling and tool integration framework. This allows the models to call external tools mid-reasoning, a capability OpenAI says extends their usefulness for agentic tasks in which a model must take sequential actions to complete a goal.

Availability and Pricing

Access to o3 is included for ChatGPT Plus subscribers at a usage tier, with higher usage available to Pro subscribers. API pricing for o3 is set at $10 per million input tokens and $40 per million output tokens at launch. o4-mini is priced at $1.10 per million input tokens and $4.40 per million output tokens through the API, positioning it as a cost-effective option for high-volume applications.

The models are accessible through the ChatGPT interface and the OpenAI API. Enterprise and Education plan availability was also confirmed at launch.

Competitive Context

The release comes amid active competition in the reasoning model segment. Google DeepMind has released reasoning-capable versions of its Gemini 2.0 series, and Anthropic has indicated its Claude model line incorporates extended thinking features. Meta's Llama series and several open-weight models from other research groups have also targeted reasoning task benchmarks in recent months.

OpenAI is scheduled to continue its stated "12 days of releases" cadence, with additional model announcements and feature updates expected in the days following the o3 and o4-mini launch.