OpenAI o1 pro

What it is

o1 pro is a version of OpenAI's o1 model that uses more compute to think harder and give better answers to the hardest problems^[1]. The idea behind models of this kind is that a model can work through intermediate steps before it commits to an answer, and that spending more of that work helps on questions a single pass does not settle. o1 pro is the variant in which that work is dialled up.

Launch

OpenAI introduced it on 5 December 2024 under the name "o1 pro mode", as part of a new ChatGPT Pro plan priced at $200 a month^[1]. Inside ChatGPT it appeared as an option users picked in the model picker^[1]. In launch form it was therefore both a model and a product decision: the more compute-heavy way of running o1 was sold as part of a $200-a-month subscription rather than as a free-standing release. The name itself described a mode — a way of running o1 with more thinking — rather than a separate model family.

What OpenAI reported

OpenAI said external expert testers found that o1 pro mode produced more reliably accurate and comprehensive responses, especially in data science, programming and case law analysis^[5]. On challenging benchmarks in maths, science and coding it performed better than both o1 and o1-preview^[5].

Reliability was the part OpenAI stressed. It evaluated the model with a stricter setting, in which a question counted as solved only if the answer was right consistently^[5]. That changes what a score means. A model that lands on the correct answer occasionally can pass an ordinary test; it does not pass one that requires the answer to be right every time. Under this measure, what counts is consistency rather than a single success, and that is the claim the extra test-time compute is meant to support.

Through the API

The model reached the API on 19 March 2025, at $150 per million input tokens and $600 per million output tokens^[2][3]. The API specification gives a 200,000-token context window, up to 100,000 output tokens, a knowledge cutoff of 1 October 2023, and both text and image input^[2]. The snapshot was named o1-pro-2025-03-19, and OpenAI scheduled o1-pro for shutdown on 23 October 2026^[2][4].

Pricing per million tokens, separately for input and output, lets a developer estimate cost from the length of what is sent and what comes back. Reasoning models produce long outputs, so the ceiling of 100,000 output tokens is a working part of the specification rather than a detail: the intermediate work and the final answer both have to fit. The context window sets the other limit, on how much material can be handed to the model at once, and the stated knowledge cutoff marks where its built-in knowledge ends.

Where it sits

o1 pro records a stage in how reasoning models were sold. It is a version of o1 with more compute spent at answer time, first offered as part of a subscription tier rather than on its own. The later API availability opened the same model to developers, and the scheduled deprecation date shows that even a flagship variant carries a defined lifetime.