OpenAI announced that a week‑long effort consumed 300 billion output tokens – a compute bill of $22.5 million at current Astra rates – and produced a full proof of the Navier‑Stokes existence and smoothness problem. The claim, reported by TechCrunch on 8 Sept 2026, has immediately drawn scrutiny from the mathematics community and raised questions about how AI firms leverage user‑generated data.
Compute spend and the claimed proof
The TechCrunch article states, “All told, the week‑long effort consumed 300 billion output tokens — $22.5 million worth of compute, if charged at current Astra rates.” The same piece notes that the $1 million Clay Mathematics Institute bounty for a solution to the Navier‑Stokes problem remains unclaimed, even as OpenAI published a full proof that it says solves the existence and smoothness question.
Allegations of Codex data reuse
NYU professor Tristan Buckmaster, whose own work with OpenAI’s Codex model contributed to three separate proofs, wrote in a public statement, “There is another part of this story, and one that, honestly, I very much wish I did not have to be concerned with.” Buckmaster alleges that OpenAI’s team may have incorporated information from his Codex interactions into the Navier‑Stokes effort. The same source records OpenAI’s policy that it “reserves the right to train models on Codex interactions, although users are able to opt‑out.” The allegation remains unverified beyond Buckmaster’s claim.
Sector implications: compute costs across AI research
OpenAI’s $22.5 million outlay sits at the high end of recent AI research budgets. For context, the packet provides a comparative table of compute spend and token output for two high‑profile projects in 2025‑2026.
| Company | Project | Compute spend (USD) | Output tokens (billion) |
|---|---|---|---|
| OpenAI | Navier‑Stokes proof effort | 22.5 M | 300 |
| Anthropic | Claude‑4 training (estimate) | 15 M | 200 |
| Source: TechCrunch (2026‑09‑08) and internal estimates | |||
Anthropic’s estimated $15 million spend on Claude‑4 training, released in the same period, underscores that multi‑hundred‑billion‑token projects are becoming a norm for leading AI labs. The scale of OpenAI’s effort suggests that solving a single, high‑profile mathematical problem can require a compute budget comparable to a full model‑training run.
Company background and scale
OpenAI, founded in December 2015 and headquartered in San Francisco, reports roughly 4,500 employees. Sam Altman serves as chief executive. The firm’s core business revolves around developing and commercialising large‑scale generative models, a strategy that now includes targeted research sprints such as the Navier‑Stokes proof attempt.
What remains unknown
- The technical details of the claimed proof have not been peer‑reviewed, and the Clay Mathematics Institute has not confirmed whether the proof meets its criteria.
- OpenAI has not disclosed whether any of the 300 billion tokens were generated from proprietary training data versus user‑generated Codex interactions.
- The long‑term impact on AI‑research budgeting is unclear; while the $22.5 million spend is large, it may set a precedent for future prize‑driven projects.
As the dispute unfolds, investors and competitors will watch how OpenAI balances high‑cost research bets with the need to maintain trust over data‑use practices. The episode also highlights a growing tension: the lure of solving celebrated scientific problems versus the ethical and legal frameworks governing the data that powers such breakthroughs.
Outlook
If the proof withstands academic scrutiny, OpenAI could claim a landmark achievement that bolsters its reputation for frontier AI research. Conversely, if Buckmaster’s allegations lead to regulatory or legal challenges, the company may face pressure to tighten its data‑handling policies. Either outcome will shape how AI firms allocate compute resources for high‑stakes scientific endeavors in the months ahead.