OpenAI announced a Navier-Stokes breakthrough using Astra-next and $40M in compute — then a mathematician accused them of stealing his solution path from Codex

OpenAI announced yesterday that Astra-next solved the Navier-Stokes existence and smoothness problem — one of the seven Millennium Prize Problems with a standing $1M award since 2000. The run took 88 hours, roughly 10,000 agents, 130 billion tokens, and cost over $40 million. It's the kind of result that should dominate AI news for a week.

Instead, within hours, mathematician Tristan Buckmaster publicly accused OpenAI of stealing his solution approach. Buckmaster says he had been working on the same problem with a co-author who works at Anthropic, uploading drafts to Codex for formatting help. An OpenAI researcher allegedly contacted him after learning about his progress, pressured him to remove the Anthropic co-author, and threatened him when he refused. Days later, OpenAI announced its own breakthrough using what Buckmaster describes as the same unusual solution path he had documented in Codex.

Buckmaster asked OpenAI whether the model had accessed his user data. He says they told him it hadn't looked up his files, but when he asked specifically about training data, he got no answer. OpenAI has denied the allegations but hasn't addressed the training question publicly.

Terence Tao commented on Mathstodon that "even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential." He noted that incentives now point toward not sharing research directions publicly at all.

What the Codex upload timeline means

Codex has been part of OpenAI's developer tooling since 2021, rebranded under GitHub Copilot and later folded into ChatGPT's code interpreter modes. Users upload notebooks, LaTeX drafts, proofs-in-progress. The terms of service have always said user data won't be used for training without permission, but the line between "looking up" a file for context and incorporating it into a training corpus is not cleanly defined in the product.

If Buckmaster uploaded working drafts with partial proofs and novel approaches, those files sat in OpenAI's infrastructure. Whether they were indexed, cached, or fed into fine-tuning runs for Astra-next is the question no one has answered. The timing is suspicious: Buckmaster says the OpenAI researcher contacted him shortly after his progress became known in academic circles, and OpenAI's announcement followed soon after.

The technical detail that makes this credible is "the same unusual solution path." Navier-Stokes has been attacked from dozens of angles over 25 years. If two teams independently land on a standard approach, that's normal. If both use a novel, non-obvious method within weeks of each other — and one team had access to the other's drafts — the coincidence is harder to explain.

The training data question OpenAI won't answer

OpenAI's public response has been to deny accessing Buckmaster's files for the Astra-next run. But when Buckmaster asked about training, the conversation reportedly went silent. That distinction matters. If Codex user uploads are excluded from model training entirely, OpenAI should say so clearly. If they are included under certain conditions (aggregated, anonymized, opted-in), that should also be documented. The silence suggests the answer is "it depends" or "we don't know."

This is not an edge case. Thousands of researchers use ChatGPT, Codex, and Claude to draft papers, debug proofs, and format LaTeX. If any of that material ends up in training corpora — even indirectly, through RLHF feedback or fine-tuning on "high-quality academic reasoning examples" — the boundary between tool and collaborator collapses. A user who uploads a novel proof expects the tool to help them finish it, not to train a competitor model that publishes first.

The Anthropic angle adds another layer. Buckmaster's co-author works at Anthropic. If the OpenAI researcher was aware of that and explicitly pressured Buckmaster to remove him, it suggests internal awareness that the project had competitive sensitivity. Why would that matter unless OpenAI planned to publish on the same problem?

What this means for production AI tooling

We deploy agents that handle client data daily. The contracts always specify: client uploads stay in tenant-isolated storage, are not used for training, and are deleted on request. That's table stakes. But the Buckmaster case shows how easily that promise can blur when the product itself is a frontier model that improves through exposure to novel reasoning.

If you're using Claude, GPT, or any other frontier model in your research or business workflow, the training-data question is now front and center. Upload a novel sales script, a new pricing algorithm, a competitive analysis — does it get indexed? Does it inform the next fine-tuning run? The terms of service rarely spell this out in detail, and when pressed, the labs often punt to "we aggregate and anonymize" without defining the retention window or exclusion guarantees.

The OpenAI announcement was supposed to be a showcase for Astra-next's reasoning depth. Instead, it's a case study in why data provenance matters more than benchmark scores. The Navier-Stokes result might be legitimate. It might also be the first high-profile example of a frontier model trained on a user's uploaded work, then deployed to beat them to publication. Until OpenAI clarifies what happened to Buckmaster's Codex drafts, the second possibility will shadow the first.

Tao's observation is the one that sticks: the incentive structure now punishes open collaboration. If sharing progress publicly triggers an AI-powered race to flatten your work before you finish, researchers will stop sharing. That's not a frontier-AI success story. That's a regression to closed, defensive research cultures — exactly what the AI labs claimed they wanted to prevent.