Ten unpublished math problems. One AI sprint.
Meet the First Proof attempts.

What happened?

In February 2026, OpenAI published its internal model's attempts at the ten First Proof problems, a challenge built to test whether AI can produce checkable proofs in specialised research fields OpenAI post ↗. Based on expert feedback, OpenAI said at least five attempts, for problems 4, 5, 6, 9 and 10, had a high chance of being correct OpenAI post ↗.

First Proof was created by eleven prominent mathematicians, including Fields Medal winner Martin Hairer, who each contributed a lemma from their own unpublished work Scientific American ↗. A lemma is a smaller theorem used as a step toward a bigger result. Because the answers were not online, an AI could not simply look them up. The problems were released on February 5, 2026, and the organisers released their own solutions on February 13 First Proof site ↗.

OpenAI shared its attempts on February 14 and explained its process in a post dated February 20 OpenAI post ↗. It first believed its answer to problem 2 was likely correct, but after official commentary and community analysis it now believes that answer is wrong OpenAI post ↗. Chief scientist Jakub Pachocki had earlier said six of ten had a high chance of being correct, and mathematicians pointed to possible gaps in at least one of those Scientific American ↗.

What are the three pieces of the story?

How did AI help?

OpenAI says the model ran with limited human supervision OpenAI post ↗. People still played a part: researchers sometimes suggested retrying strategies that had worked before, asked for proofs to be expanded after expert feedback, used ChatGPT to help with checking and formatting, and for some problems chose the best of several attempts OpenAI post ↗. The mathematical arguments themselves are presented as the model's work.

Figures are OpenAI's own assessment from its February 20, 2026 post OpenAI post ↗.

OpenAI was open about the limits. It called the effort a fast sprint and said its process was not as clean as a properly controlled evaluation OpenAI post ↗. That matters because First Proof organiser Lauren Williams asked how to judge how much is human and how much is AI once humans are involved Scientific American ↗. A Google team running a different system reported that its agent solved 6 of the 10 problems on its own by majority expert assessment, which shows that the scores of different labs used different rules arXiv preprint ↗.

Which fields could this affect?

The immediate value is a public record of how an AI model performs on fresh research problems; wider uses are possible later, and these connections are our assessment.

What has been checked?

The evidence is a set of published proof attempts, the company's own assessment, organiser solutions and news reporting, not a peer reviewed evaluation. Leapscope reviewed these sources; we did not repeat the experiments.

Evidence status: Proof attempts. Stage: Claim. Proof attempts shared publicly; independent checking is not complete.

From sprint to fair test

1. Finish the expert checks.Specialists in each field need to confirm or reject the remaining attempts.
2. Run a controlled test.OpenAI says it hopes to discuss a more rigorous evaluation framework with the organisers OpenAI post ↗.
3. Compare under the same rules.Results from different labs only mean much when the human help allowed is the same.

This is our suggested way to follow the story, not a promised timetable.

Can I use it today?

You can read OpenAI's post and its preprint of all ten attempts, and the organisers' solutions and comments OpenAI post ↗ Organiser solutions ↗. The model that wrote the attempts is internal and not available to the public OpenAI post ↗.

A few things you might be wondering