OpenAI has released o3-mini. Here's what to know.

• OpenAI has released o3-mini, a small "reasoning" model designed for STEM, logic and coding tasks.

• It's available for free via ChatGPT and costs the same as o1-mini on the API.

• It may offer better results than o1 for technical applications.

OpenAI has launched a mini version of its next-generation o3 "reasoning" model via ChatGPT and API. Designed for technical tasks, the company says it excels at scientific, engineering, math and coding queries.

The move comes days after Chinese rival DeepSeek sent stock markets reeling with its low cost R1 model. Like R1, ChatGPT users can access o3-mini for free. The model can also search the internet and reference its sources.

In a Reddit AMA yesterday, CEO Sam Altman hinted that the full o3 model should drop in the next couple of months.

It's part of a wider company focus on scientific reasoning, which Altman says will ultimately produce the greatest benefits for society.

"The most important impact [of AGI], in my opinoin, will be accelerating the rate of scientific discovery, which i believe is what contributes most to improving quality of life."

As well as o3, he and his colleagues said the firm was working to improve its image generation and agentic offerings, and investigate robotic applications. If they're right, a future where agentic robots perform everyday tasks might not be that far off.

Wider access and more capabilities

o3-mini has launched with more capabilities than its predecessors, enabling it to compete more directly with DeepSeek R1, which offered API and free chat access on release.

The model can be used with ChatGPT tools like Canvas, and it can search the web unlike o1, OpenAI's OG reasoning model. But it doesn't yet have vision skills.

It is, however, designed to be more developer-friendly than its predecessors, offering function calling, structured outputs and streaming out-of-the-box.

CPO Kevin Weill said during yesterday's AMA that the firm ultimately wants to offer a more comprehensive service for all its models and tools, several of which don't currently work with each other.

"We have a lot of good tools that were built somewhat independently—goal is to get them in your hands as fast as possible. The next step is integrating all of this ... All the tools need to work together better. And btw all the models need full tool use (o series models can't yet use all tools), that will happen too."

Speed and performance

o3-mini is reportedly 24% faster than o1-mini, and should reply more accurately. But users can toy with speed to optimize performance for their particular use case.

Developers can choose between "low", "medium" and "high" effort versions of the model, essentially changing how quickly the model "thinks" about its answer. More thinking time should deliver better answers to more complex questions. But responsese will inevitably take longer.

The "high" version produces impressive results for those happy to wait. OpenAI's benchmarking results show it outperforms even the full version of o1 on a number of math, science and coding tests.

Does this mean o1 is already redundant? Apparently not, according to OpenAI, which says o1 will function as a generalist to o3-mini's STEM specialist. Despite this, o3-mini will replace OpenAI o1-mini in the ChatGPT model picker.

Benchmarking tests only provide a limited picture of performance, so good results may not translate into high performance for your own use case.

The model won't be appropriate for plenty of use cases. It's overkill for genuinely simple requests, and can treat throwaway user responses as full-on prompts. As you can see below, it spent 10 seconds reasoning through its response to a simple "cool thanks" user message:

Access and pricing

In ChatGPT, paid users can select o3-mini in the model picker. Free users need to press the "reason" button when sending a request to access o3-mini. ChatGPT will use the "medium" version of the model as standard, with paid accounts also able to choose "high."

Plus and Teams users get 150 messages per day, while Pro subscribers have unlimited access to o3-mini. OpenAI doesn't specify how many o3-mini messages free users will get, but The Verge expects similar limits to GPT-4o.

Enterprise users should be able to access o3-mini sometime this month.

The model is available via the OpenAI API platform, but not to everyone just yet. The company is rolling it out to "select" accounts at Tier 3 and above (who've spent at least $100 via the OpenAI platform).

Once access is granted, developers will be able use the model via the Chat Completions, Assistants and Batch APIs.

o3-mini is available at the same price as o1-mini:

• Input cost: $1.10 per million tokens

• Cached input: $0.55 per million tokens

• Output: $4.40 per million tokens

If you're using o1 and technical, coding or scientific applications, it might be worth switching as o3-mini is much cheaper. For comparison, o1 currently costs $15 per million input tokens.

You can cut costs by 50% for both models if you use OpenAI's Batch API.