OpenAI releases o3-pro

Background

Reasoning models differ from ordinary chat models in how they answer: instead of producing a reply in one pass, they work through a problem step by step before responding, spending extra computation in exchange for accuracy. Above that baseline, OpenAI also sells higher-compute variants under a "pro" name — the same kind of model given more room to think, aimed at harder problems.

That top tier had a placeholder in public view before o3-pro arrived. When OpenAI introduced o3 and o4-mini, the launch post said that o3-pro would follow "in a few weeks" and would ship with full tool support, and that ChatGPT Pro subscribers would keep using o1-pro until then^[3].

What was released

On 10 June 2025, OpenAI released o3-pro^[1]. It took over from o1 pro as the company's highest-compute reasoning option^[1] — the model in the lineup that is given the most computation to work with when generating an answer.

OpenAI positioned it for questions where reliability matters more than speed^[1]. That is a fairly precise statement about trade-offs rather than a claim about general capability. A model that spends longer reasoning is not the right choice for every request: for routine drafting, summarising or quick lookups, waiting longer buys little. The case for a model like o3-pro is the opposite kind of task — the one where a wrong answer costs more than the extra wait, so it is worth allocating more computation to the answer before it is given.

Access and pricing

At launch, o3-pro was offered to ChatGPT Pro subscribers and through the API^[3]. The subscription route meant the model appeared inside an existing paid plan rather than as a separate product, while the API route let developers call it directly from their own applications.

On the API, the published prices were $20 per million input tokens and $80 per million output tokens^[2]. Tokens are the units that models read and write in — roughly word fragments, with longer prompts and longer answers costing more. Listing input and output separately is standard practice, and the gap between the two numbers reflects the fact that generating text is the more expensive operation. The practical consequence for a developer is that cost scales with use: a pipeline that sends long documents in and gets long reports out pays far more per call than one that asks short questions.

Why it matters

The release shows how reasoning models are being packaged. Rather than one model per generation, OpenAI has been shipping a family: standard reasoning models, smaller and cheaper variants, and a top tier that consumes the most compute. o3-pro sits at the top of that stack, and the fact that it was announced ahead of time in the o3 and o4-mini launch post — with a promise of full tool support and a stopgap arrangement for Pro users still on o1-pro^[3] — shows that the tier is planned as part of the lineup rather than added afterwards.

For users, the meaningful change is a choice that is now explicit. The question is no longer only which model is most capable, but which one matches the job: a fast, inexpensive model for high-volume work, or a slower, more expensive one for the cases where being wrong is the expensive outcome. Pricing is part of that decision, and per-token rates in the tens of dollars per million tokens set the terms on which developers weigh a heavier model against a lighter one^[2].