OpenAI takes back $1m of Mathathon credits on the day 771 mathematicians publish their letter

Round two of the Mathathon ends in an arXiv preprint, a video presentation and an auto-formalised solution ^[19]. The letter's charge is about what happens to that preprint next. Research mathematicians are the ones who verify, spread and correct the results AI companies announce, and that labour "goes uncompensated, uncredited, and unacknowledged", the letter says ^[5]. It concludes that AI companies are engaging in research misconduct ^[6].

The letter asks the organisers to suspend the event and gives five grounds, the first of which calls the work slop mathematics, produced fast and left for other people to check later ^[4]. Signing is limited to the Caltech community and mathematics researchers at PhD level or above ^[3].

Dan Roberts, an OpenAI research lead, wrote on X that the company recognises the rapid progress of AI in mathematics is disruptive, and said OpenAI wants to engage the maths community on how to integrate the technology ^[8]. Truman Dickerson reported the withdrawal for Business Insider ^[9].

A hundred teams at roughly $20,000 each needs $2m of credits, which is what the two sponsors together had put up ^[1]. With one of them gone, the committed pool is $1m, or $10,000 a team on the announced team count, half the allocation described on 4 September ^[2]. The organisers had over a thousand applications by 10 September, from undergraduates through to faculty, ten applicants for every slot ^[20][3]. Their supporter list also includes DARPA expMath and Cognition ^[21].

Anthropic spent the same fortnight courting the field. TNW reported on 6 September that it used Claude to formalise Fermat's Last Theorem in Lean in 11 days, and it has since widened free access and research credits for formalisation work ^[12]. The auto-formalisation requirement in round two points at the same category of work ^[19]. Business Insider said Anthropic did not immediately answer its request for comment ^[11].

The Fields Medallists' declaration asks for no cancellation and never names the Mathathon; it is a statement about benchmarks ^[17]. It says solving problems is only a tool and a proxy, and that the primary goal is conceptual understanding and insight ^[15]. Mass production of true or false statements at increasing pace "could destroy fertile ground", the signatories write ^[16]. The Economist put the number of signatories at 24; the declaration itself lists 25, with medals awarded between 1978 and 2026, and Terence Tao described himself as one of 25 initial signatories ^[14].

The five named organisers are students, and they published a response the same day as the letter, conceding that several of the concerns are valid and that their original description did not address them ^[22]. They added the arXiv requirement the letter had asked for, expanded the educational track, and cut a line from the website asking what the role of a mathematician is when AI can solve conjectures faster ^[23].

The line in the letter that travels outside mathematics is the one about per-seat cost. The test it suggests for any rollout is whether producing the output and checking it sit in the same budget. When they do, bad output shows up in your own numbers. When they do not, the pilot is running on a subsidy from people who never agreed to fund it, and it lasts as long as the credits do. No reasonable future model of mathematical research gives every research mathematician $20,000 in AI credits to prove a result, the letter says ^[7].