OpenAI's First Proof submissions and the human scaffolding around them

What OpenAI actually submitted

On February 14, 2026, OpenAI published proof attempts for all 10 problems in First Proof, a challenge built to test whether AI can produce correct, checkable proofs on domain-specific research problems rather than short-answer or competition math.

Based on expert feedback, OpenAI says at least five attempts — problems 4, 5, 6, 9, and 10 — have a high chance of being correct, with several others still under review. Notably, the company revised its own claim: it initially believed problem 2 was likely correct, then concluded it was incorrect after the official commentary and community analysis.

That public correction is the honest detail. It shows correctness here is not self-certifying; it depends on a review process OpenAI does not control, and OpenAI's own confidence was wrong on at least one problem.

The interaction layer did real work

OpenAI is unusually candid that these proofs came out of a supervised, iterative process — not an autonomous run. The description of the workflow is where the substance lives.

By its own account, OpenAI sometimes suggested retry strategies that had worked in earlier attempts, asked the model to expand or clarify parts of a proof after receiving expert feedback, and facilitated a back-and-forth between the model and ChatGPT for verification, formatting, and style. For some problems, it presented the best of a few attempts, selected by human judgment.

That is a substantial frontend around the model: a human steering strategy, an external verifier model, and manual selection at the end. The preprint even adds an appendix of prompt patterns meant to simulate those manual interactions — an acknowledgment that the prompting itself is part of the result.

A moving target, not a fixed evaluation

The model wasn't held constant during the exercise. Researcher James R. Lee describes trying the problems over a weekend while the model was still in training and watching its capability grow.

Already it was able to solve two of the problems (#9 and #10). As it trained, it became increasingly capable, eventually solving–in our estimation–at least three more. ... It's pretty incredible to watch a model get tangibly smarter day by day.Montana Labs

That framing is exciting, but it also complicates the numbers. Attempts submitted at different points came from different checkpoints of a model being actively trained. OpenAI concedes the point directly: "This was a fast sprint, and our process was not as clean as we would like in a properly controlled evaluation," and it says it wants to work with the First Proof organizers on a more rigorous framework.

Why the process disclosure matters more than the count

OpenAI positions First Proof as evidence that novel frontier research — sustaining long reasoning chains, choosing abstractions, handling ambiguous problem statements, producing arguments that survive expert scrutiny — is a better probe of capability than benchmarks. It's a continuation of a stated arc: IMO gold-medal performance (35/42) in July 2025, the GPT-5 science case studies in November 2025, and a GPT-5.2 gluon-amplitude result later verified by authors.

The specific implication for anyone building on research-grade reasoning: read the workflow, not the headline. Five likely-correct proofs came with retry coaching, a second model as verifier, expert-driven clarification rounds, and human best-of selection — and one confident claim still turned out wrong. For applied teams, that pattern is the deliverable to study. The reproducible artifact isn't a benchmark number; it's the appendix of prompt patterns and the verification loop that surrounded the model.