Our First Proof submissions

OpenAI just published a batch of proof attempts from one of its models tackling problems from something called the First Proof challenge, a set of expert-level math questions designed to push past the usual benchmark fare. Instead of another leaderboard score, the company is showing its work: actual attempted proofs, warts and all, for anyone to read and pick apart.

That's a meaningfully different move than the typical AI math announcement. Most of the industry loves to cite a percentage on some competition math benchmark and move on. Here, OpenAI is putting the raw output in front of mathematicians and letting them judge whether the reasoning actually holds up, step by step, rather than just trusting a final numeric answer.

The stakes matter because proof-writing is a different skill than answer-finding. A model can guess the right number on a calculus problem through pattern matching, but constructing a rigorous, gap-free argument is the kind of task that exposes whether a system truly understands the logical structure underneath, or is just very good at sounding like it does.

OpenAI isn't claiming these attempts are flawless research breakthroughs. The framing is closer to a transparency exercise: here's where the reasoning holds, here's where it wobbles, judge for yourself. For a field that's been accused of overselling benchmark wins, publishing the messy middle is at least a more honest way to make the case that these models are creeping toward genuine mathematical reasoning.