Terence Tao: OpenAI Chased a Math Prize to Beat Anthropic

The mathematician Terence Tao told The Observer, in an interview published on 23 October 2026, that OpenAI’s push on a Millennium Prize problem “became code red priority number one” after a rumour that Anthropic was close to one, citing his contacts at OpenAI. The account comes from the summary of his views on AI that Tao keeps on his own website, which quotes the interview.

In the same interview, as quoted there, Tao said the labs are “racing towards the most powerful model” and must “train and benchmark and give the AI lots of tests”, so that “maths problems are almost the only thing left for them to train on, so suddenly all the big prizes and goals in our profession are benchmarks for them.”

Tao, a Fields medallist, gives this as his reading of events; OpenAI and Anthropic did not respond to these claims in any text we read. This piece sets his words next to the AI math releases we have reported this month.

Business Pill · THE AI AUDIT PROBLEM

A one-minute explainer of the AI audit problem: when producing work becomes cheaper than checking it. It teaches the general idea only and says nothing about any organisation in this story.

What Tao Said

Tao’s site quotes his complaint about the output itself: “And as a byproduct, they’re generating these terrible proofs. Great results, but they don’t care how elegant the solutions are, or who understands them, or what to do next.”

The summary gives his account of the trigger, again citing his contacts at OpenAI: tens of millions of dollars went in, a swarm of ten thousand agents was aimed at Navier–Stokes, and signups to the premium service were suspended for days to redirect the compute. Tao is quoted as saying: “They really, really wanted to beat Anthropic. It’s insane.”

He described what followed as a disposal problem. The company had “a bunch of very badly polished solutions”, he said, and “when you generate toxic waste, you have no good options. You can sit on it forever, or you can just dump it out all at once, or you can dole it out in drabs, or you can pay people to process it.”

According to the same summary, Anthropic chose the last option, paying mathematicians to write up results from its own trove, which Tao says raises a different problem: “they get to pick who writes up the paper.”

Earlier, to The New York Times on 16 October, he called the Navier–Stokes proof “technically readable” but said “no expert I’ve talked to has actually understood the whole argument”. The summary says that before the October release he had called the coming results “toxic waste”, and afterwards “maybe treated waste.”

What Was Posted on X

The X account @haider1 posted on 24 October a paraphrase of Tao’s interview, with a screenshot of the article’s headline. The wording in the post is the poster’s summary, not Tao’s; the quotations in this piece come from the summary on Tao’s own website.

Hesamation posts that more than 1,000 mathematicians have joined the Association for Human Mathematics, committing not to publish AI-generated math papers or work with frontier labs, and that 654 of them also pledged not to use AI in their research. The counts are the post’s and were not independently verified.

The October Math Releases in Our Reporting

Each row comes from one of our earlier pieces, in the words of the company or researcher that reported it. The rows measure different things, so we list them and do not add them up.

By our arithmetic, the host’s $40 million estimate for one run is about 2.7 times the $15 million the three backers committed to help mathematicians. The estimate multiplies public token counts by public list prices and, as our piece on it noted, is not what OpenAI spent.

Connecting the Coverage

On 6 October, OpenAI published “Sharing AI progress in mathematics” and a catalogue of 722 manuscripts, as we reported in OpenAI’s Internal Model Generated 722 Math Papers. On 9 October, Tao told a public lecture at Caltech that “Further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics,” as we reported in Terence Tao: Optimizing Only for Problem-Solving Harms Math.

Our earlier pieces on the Navier–Stokes claim covered OpenAI’s math advisory group and the token count behind the run. On 15 October, the Simons Foundation said it, Anthropic and OpenAI would commit $15 million, with the foundation managing and awarding the money, as we reported in OpenAI and Anthropic Back Simons’ $15M AI-for-Math Push.

What’s new in this interview: Tao’s account of why the race happened, the rumour about Anthropic, and the claim that Anthropic is paying mathematicians to write up its own results.

The Structural Read

As we read it, Tao’s complaint is about who carries the follow-up work: he says the company had “a bunch of very badly polished solutions”, and lists the options as sitting on them, dumping them, doling them out or paying people to process them.

As we read it, the money figures in our reporting sit beside his account: a Latent Space host estimated one run at about $40 million at public prices, while the Simons Foundation, Anthropic and OpenAI committed $15 million to help mathematicians navigate AI.

As we read it, the claims about motives rest on Tao’s contacts at OpenAI, and neither OpenAI nor Anthropic confirmed them in any text we read.

Three Implications

FOR MATHEMATICIANS Tao says Anthropic pays mathematicians to write up results from its own trove, and that “they get to pick who writes up the paper.”

FOR AI LABS OpenAI’s repository says each result used, on average, three hours of ChatGPT Pro thinking compute, and the Latent Space host put the Navier–Stokes run at 10,000 agents over 88 hours.

FOR FUNDERS The Simons Foundation says it will manage and award the $15 million, including through an open call for proposals.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework puts it as “The Verifier Moat — built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”

As we read it, Tao describes mathematics as one of those reward signals: problems with checkable answers that labs use to train and benchmark their models. The framework was published in July 2026 and does not discuss these events.

What Is Not Established

In the texts we read, OpenAI has not confirmed the code-red account, the spending or the suspended signups, and Anthropic has not confirmed that it pays mathematicians to write up its results. Tao attributes the account of OpenAI’s motives to his contacts there.

We did not read The Observer’s article directly. The quotations above are as reproduced in the summary on Tao’s own website, last updated 23 October 2026, which names the article and its author, James Tapper.

The Bottom Line

Tao told The Observer that the big prizes in mathematics have become benchmarks for AI labs, and that OpenAI’s Navier–Stokes push followed a rumour about Anthropic, according to his contacts. Our October reporting lists 722 OpenAI manuscripts at release (719 after three were withdrawn), a host’s $40 million estimate for one run and a $15 million fund to help mathematicians.