OpenAI published 722 math manuscripts. Mathematicians want to see the work.
OpenAI posted 722 AI-written mathematical manuscripts from an unreleased model, claiming most were generated from a single prompt, covering number theory, algebraic geometry, and more. The coverage split cleanly between outlets treating the volume as the story and outlets treating the absence of peer review as the story. Decrypt was the only outlet to name the verification problem in its headline. The number that matters isn't 722. It's zero: the number of independent mathematicians who have confirmed the results.
OpenAI secret AI model solves hundreds of open math problems in one prompt
4 sources · hover a dot to see coverage
What happened
OpenAI released 722 AI-generated mathematical manuscripts attributed to an unreleased model, claiming the bulk of the output came from a single prompt. The release spans fields including number theory and algebraic geometry, and follows by roughly one month a separate claim that an OpenAI model solved the Navier-Stokes problem, a longstanding open problem in fluid dynamics. Fortune reported the total as more than 370 outstanding math challenges addressed, a number that differs from the 722 manuscript count, suggesting the two figures measure different things. Mathematicians' responses ranged from celebration to sharp criticism, with some describing the effort as an assault on their field. The model used has not been publicly released, and the manuscripts have not undergone peer review. That is the wire version. Four outlets have it. What separates the coverage is how seriously each one treats the word 'claims.'
Decrypt named the verification problem; everyone else buried it
Decrypt's headline reads 'Mathematicians Want Receipts,' the only outlet of the four to put the unverified status in the lede position. Its excerpt specifies that 'most came from a single prompt' and that the claims are 'unverified,' framing the story as a credibility question rather than an achievement announcement. The other three outlets led with the scale of the output or the drama of the reception. That editorial choice has a concrete consequence: a reader who saw only the Fortune or Quartz headline would have no reason to ask whether any of the 722 manuscripts had been checked by anyone outside OpenAI.
Fortune counted 370; Decrypt counted 722. The discrepancy went unaddressed.
Fortune's piece refers to 'more than 370 outstanding math challenges,' while Decrypt and Quartz both work from the 722 manuscript figure. None of the four outlets explained the gap between those numbers, and none flagged it as a discrepancy worth resolving. The two figures likely measure different things, possibly manuscripts versus distinct problems, but that distinction matters enormously for evaluating the claim. A reader moving between outlets would encounter two different numbers describing the same release with no reconciliation offered.
The Economist treated 'claims' as the operative word and stopped there
The Economist's headline uses 'claims to have solved' where Fortune writes 'publishes solutions' and Quartz writes 'solved.' That single word does real epistemic work, but the Economist excerpt available here offers nothing beyond the headline, so it is not possible to assess how deeply the piece interrogates the claim. The headline posture is the most cautious of the four, but caution in a headline without follow-through in the reporting is just hedging.
What one side told you that the other didn't
722 manuscripts, zero peer reviews. Three outlets skipped the second number.
None of the four outlets reported that any of the 722 manuscripts had been independently verified by mathematicians outside OpenAI. Decrypt named the absence explicitly. Fortune, Quartz, and the Economist did not. The model that produced the manuscripts has not been publicly released, which means external verification is structurally impossible until OpenAI allows it. Framing a release of unverifiable outputs as 'solutions' is a choice. Three of four outlets made it without flagging it.
Fortune said 370. Decrypt said 722. Nobody explained why.
Fortune reported 'more than 370 outstanding math challenges.' Decrypt and Quartz reported 722 manuscripts. The gap between those figures, roughly 350, is large enough to change the story's scale entirely, and no outlet reconciled them. The most likely explanation is that the two numbers measure different things, problems versus manuscripts, but that distinction is precisely what a reader needs to evaluate OpenAI's claim. The discrepancy sat in plain sight across four pieces of coverage and none of them named it.
The Navier-Stokes claim from a month ago got one sentence total.
Quartz noted that the current release 'came a month after OpenAI's AI solved the Navier-Stokes problem.' That prior claim, itself extraordinary and itself unverified by the broader mathematical community at the time of this release, received one subordinate clause across the four outlets. If the Navier-Stokes result was not independently confirmed in the intervening month, that context changes how the 722-manuscript release should be read. None of the four outlets reported on the verification status of the earlier claim.
What to watch
The manuscripts are public, which means independent mathematicians can begin checking them now. If a credible research group publishes a formal assessment of even a subset of the 722 papers within the next 30 days, watch whether outlets that led with the achievement frame update their coverage or treat the verification as a separate story. If the assessment is critical, the gap between 'publishes solutions' and 'claims to have solved' will become the story.
See how outlets across the political spectrum framed this differently — and what each side left out.