OpenAI launches o3 and o4-mini, reasoning models with tool access

OpenAI has made its o3 and o4-mini models available to users today, introducing a new generation of reasoning systems: models designed to spend more compute on complex problems before responding. Their main innovation is not just stronger reasoning, but the ability to decide for themselves when to use ChatGPT’s tools.

That includes searching for information on the web, analyzing files and data with Python, interpreting images, and generating images. Until now, these capabilities existed separately or required more explicit instructions from the user. With o3 and o4-mini, OpenAI wants the model to chain those steps together to fulfill an entire request.

A model that combines search, code, and images

The practical difference is most apparent in questions that cannot be solved with a single piece of text. Faced with a query about projected electricity consumption in California this summer, for example, the model can search for recent public data, process it with code, produce a forecast, and present a chart with an explanation.

The model does not gain general control of the user’s computer or act outside the tools enabled in ChatGPT or through function calls in the API. But it can plan a sequence of actions within that environment: consult a source, check the result, reformulate the search if information is missing, and use the data to build an answer.

This is a significant shift for office work, analysis, and programming tasks. An assistant’s usefulness depends not only on whether it can write a good response, but also on whether it knows when it needs up-to-date data, reproducible calculations, or a careful reading of a document. The risks change too: a poorly designed chain of steps can produce an apparently solid conclusion based on weak sources or faulty assumptions. Autonomy does not replace human review, especially for financial, legal, or scientific decisions.

Thinking with images, not just describing them

OpenAI highlights another capability: o3 and o4-mini can incorporate images into their reasoning process. In simple terms, they do not merely identify what appears in a photograph or chart; they can enlarge, rotate, or transform the image while trying to solve the problem.

This could help interpret a photographed whiteboard, a technical diagram, a scanned table, or a hand-drawn sketch, even when the image is poor quality. The company calls this capability thinking with images, although users do not receive a complete transcript of the model’s internal reasoning. What matters is the verifiable result: which data it used, what calculation it performed, and which sources it consulted.

Visual integration is particularly interesting for fields where information does not arrive as clean text: engineering, data analysis, education, scientific research, or technical support. It also removes a common source of friction with today’s assistants: having to describe a chart manually before asking questions about it.

o3 prioritizes capability; o4-mini, cost and speed

OpenAI describes o3 as its most powerful reasoning model to date. The company says it outperforms o1 in programming, math, science, and visual perception, and that external evaluators found 20% fewer major errors than with o1 on difficult, real-world tasks.

According to OpenAI, o3 achieves new highs on tests including Codeforces, SWE-bench, and MMMU. Benchmarks are useful for measuring progress, but they are not the same as guaranteed reliability in day-to-day work: they typically assess narrowly defined problems with clear correctness criteria under controlled conditions.

o4-mini is the smaller, cheaper alternative. It is aimed at high volumes of queries that require reasoning, particularly in math, programming, and visual analysis. OpenAI says it achieves the best result among the evaluated models on AIME 2024 and 2025; with access to Python, it scored 99.5% on AIME 2025 in a single attempt. The company itself cautions that this figure should not be compared with models without tools, because a code interpreter substantially reduces the difficulty of part of the exam.

In the API, o3 costs $10 per million input tokens and $40 per million output tokens. o4-mini cuts those rates to $1.10 and $4.40, respectively. That difference explains why the smaller model may be more attractive to companies that need to process large numbers of documents, handle support workflows, or automate repetitive analyses.

Available in ChatGPT and for developers

Both models are arriving today on ChatGPT Plus, Pro, and Team plans; OpenAI expects to enable them for Enterprise and Edu during the coming week. They are also available through the Responses and Chat Completions APIs, with support for function calling—the mechanism that lets developers connect models to their own tools.

o3 replaces o1 in ChatGPT. For Plus and Team subscribers, OpenAI sets a limit of 50 weekly messages with o3 and 150 daily messages with o4-mini, while the Pro plan offers much broader usage subject to anti-abuse measures.

The launch shows where OpenAI is placing its next bet: not so much on a chatbot that answers better, but on one that can break down a task, find what it needs, and produce a result in different formats. Whether that promise holds will depend less on a benchmark demonstration than on whether those tool chains prove reliable, auditable, and affordable enough to use outside the lab.