OpenAI and Anthropic Fund the Checking of AI Maths

The Simons Foundation, Anthropic and OpenAI will commit $15 million in combined funding and resources to help mathematicians judge which AI-assisted results hold up, the foundation announced on 15 October 2026. OpenAI is putting in $5 million of that, The Verge and The Independent both reported the same day. Anthropic's share has not been disclosed, and the announcement does not separate cash from the value of donated tools and computing time.

The foundation will administer and award the money, including through an open call for proposals, to support workshops and initiatives that, in the words of its announcement, help researchers "evaluate recent AI-assisted results, learn and use new tools, share perspectives and shape collective standards for AI use". Anthropic and OpenAI will also provide AI tools and computing resources to researchers the foundation supports, and that part of the arrangement reaches past mathematics into basic science.

Nine days after the flood

The date sets the context. On 6 October OpenAI posted 722 mathematical manuscripts in 372 result families, produced by an unreleased internal model, as Parallax Nexus reported at the time. The company withheld the model's name, the prompts and all but ten of the reasoning summaries. The Independent reported on 15 October that the donation follows criticism from the mathematical community over that release, and The Verge noted that the money arrives just over a week after the company, in its words, dumped hundreds of preprints on the research community.

What the mathematicians were short of was not interest but capacity. Checking a proof is slower than generating one, and a result that reads as correct is not the same as a result that has been verified. The foundation has a claim on that problem: it was an early funder of the development of Lean, the proof assistant that lets a machine check an argument line by line, and which its announcement credits with making AI-assisted proofs possible in the first place.

Who gets to set the standard

That is the sentence the structure is built around, and it is also the sentence the structure strains against. The two labs whose output prompted the argument are now among the three parties paying for the forum in which the argument will be settled, even with a philanthropy holding the chequebook and running the open call. Spergel's framing concedes the risk by naming it. The honest reading is that the governance question is not answered by who administers the grant, but by whether the standards the workshops produce are ones a lab would find inconvenient.

Against that, the alternative on offer has been worse. Google DeepMind took the slower route with the same technology: the peer-reviewed account of its AlphaProof Nexus system, which solved nine Erdos problems, appeared in Science on 8 October, Science News reported, with a description of how the results were obtained and human mathematicians having vetted many of the proofs. Thomas Bloom, the mathematician who curates the Erdos problems list, told Science News that this was "quite different" from how OpenAI and Anthropic have been handling their mathematical results. A fund that pays people to do that checking is a response to a real shortage, whoever wrote the cheque.

Several things are not public. The announcement names no models, gives no figure for the compute, sets no per-lab breakdown beyond OpenAI's $5 million, and gives no date for the open call or the first workshops. Nor does it say who decides which AI-assisted results are worth a reviewer's time, which is the decision that will determine whether $15 million buys standards or a backlog.