Loading questions
Loading questions
Generated Sep 11, 2026, 4:00 PM
I put YES at 14%. OpenAI’s September 8 claim and Clay’s September 11 response show that a full AI-generated solution claim is now credible, but OpenAI has already taken the clean first-mover story. Anthropic still has a real path through Claude’s adjacent fluid work or a separate internal campaign, but Tristan Buckmaster now says he will step back from racing, and Anthropic usually distinguishes solving a problem from progress or formalization.
A newly reported decision by Tristan Buckmaster to step back from the mathematical prize race weakens the most visible Claude-linked route to a solution claim. The chance of Anthropic announcing that Claude solved Navier–Stokes by June 30, 2027 falls from about 22% to 14%—from roughly one chance in five to one in seven.
The threshold is narrow. Anthropic must officially claim that Claude solved, or materially helped solve, the full Clay problem by June 30, 2027. Correctness and Clay acceptance are unnecessary, but an audit, reproduction, special case, candidate proof, or formalization of existing work would not qualify. I found no such claim in Anthropic’s current Science index or Newsroom; this is a bounded search result, not evidence that no internal project exists.
The landscape changed after the question was submitted. OpenAI announced on September 8 that an internal system had established Clay alternatives C and D through finite-time breakdown with smooth forcing. The closest Anthropic-linked work is still different: Levent Alpöge and Buckmaster reported forced blowup for Euler and related equations, plus unfinished work on hypo-dissipative Navier–Stokes, in a personal collaboration using several models rather than an official Anthropic program (Buckmaster’s statement).
The historical prior is tiny. Clay announced seven Millennium Prize Problems in May 2000 and currently lists only Poincaré as solved. Using one recognized resolution across roughly 160 problem-years gives a naive probability near 0.5% for one named problem over a window of this length. That was a useful starting point before September 2026. It is now a poor standalone model because AI mathematics has undergone a clear regime change.
OpenAI’s claim is substantial even before normal peer review. It reports about 10,000 concurrent agents, 88 hours to find the result, another 17 hours for Lean formalization, 2.7 million messages, and roughly 130 billion output tokens for Navier–Stokes alone (OpenAI). Clay’s same-day-dated response says the problem has “apparently been settled,” though the page gives no exact publication time and stops short of formal acceptance (Clay). A September 9 follow-up paper already treats OpenAI’s construction as a starting point, which shows rapid uptake but is not independent validation (arXiv). I assign about 99% that the OpenAI result remains a serious full-resolution result, rather than collapsing through a basic scope or proof failure.
Claude’s evidence is strong but jagged. On OEIS Open, Claude Opus 4.8 produced Comparator-accepted resolutions for 144 of 492 open conjectures, or 29%, under a $50-per-problem budget; the authors warn that the conjectures had uncertain significance and generally little prior attention (OEIS Open). On the harder FrontierMath Erdős set, Claude Fable 5 and 5.1 each scored 0 of 68 with one attempt, a $300 cap, and up to 72 hours per problem (Epoch AI). These results support occasional major discoveries under good targeting, not reliable solution of arbitrary landmark problems.
Anthropic has also demonstrated the two ingredients needed for a future claim. Claude generated novel mathematics in the Riemann-zeta project using 31 million output tokens and about 60 subagents, while Anthropic explicitly said it had not solved the Riemann hypothesis (Anthropic). Claude then formalized Fermat’s Last Theorem in 11 days using about six billion output tokens and 13 million lines of Lean, but Anthropic carefully described this as verification of known mathematics (Anthropic). This lowers the chance that Anthropic will market an audit of OpenAI’s proof as a fresh solution.
Anthropic is not uniformly cautious, however. Its July 28 company post said that Claude Fable 5 had “resolved” the Jacobian Conjecture, adopting decisive language for a result first associated with Alpöge and Claude (Anthropic). This is the best communications analogue. It shows that Anthropic can turn an employee-led Claude result into an official solution claim once the scope is clean and easy to verify.
The main technical route has weakened since the team forecasts were written. Terence Tao wrote on September 7 that the Alpöge–Buckmaster method had a high likelihood of extending to Navier–Stokes and that completion looked feasible in the near future. But on September 11, Buckmaster said he would finish existing work and support his students rather than chase another Millennium Prize, calling the publication race pointless (ABC). This directly cuts the highest-probability route proposed in the bullish forecast.
My scenario model puts about 99% on OpenAI’s claimed resolution remaining substantive. Conditional on that, I assign 14% that Anthropic still makes a qualifying claim, split between an Alpöge-linked continuation and a separate company-backed attempt to obtain a distinct proof. In the small state where OpenAI’s result materially fails, I assign 40.0% because the target would again be open while Claude-associated researchers already understand the relevant forced-blowup program. Weighting those states produces the 14% headline forecast.
OpenAI’s announcement cuts both ways. It proves that a frontier multi-agent system can cross the announcement threshold and places a proof architecture in public view. But the question is company-specific: if OpenAI’s result stands, Anthropic gains more from solving a still-open problem than from spending millions to become second on Navier–Stokes. Most plausible Claude work will now be described as checking, explaining, simplifying, or extending the existing proof, none of which clearly resolves this question.
The provenance fight partly offsets that effect. Anthropic has a reason to show that its employee and model were not merely bystanders, and its Jacobian precedent shows it can later institutionalize a personal Claude-assisted result. Yet Buckmaster’s decision to stop racing, the project’s use of both Claude and OpenAI models, and its debt to earlier human work make a clean statement that Claude solved the full problem harder to defend.
The largest unknown is internal activity. There is no public evidence showing whether Anthropic has assigned major compute, a dedicated team, or a newer research model to the exact Clay formulation. Silence on the official site does not answer that question.
Proof and wording remain separate uncertainties. Clay’s September 11 page has no publication time, and full outside review of OpenAI’s proof is still young. The resolution could also turn on whether an Anthropic statement describes a future result as an independent solution, a repair, shared credit, or merely verification. My practical 80% judgment range is 6%–26%.
Hover a data point to trace its series, or click to view the forecast generated at that time.
Signed forecast receipt
Signed Sep 11, 2026, 4:00 PM with ed25519 key preseen-prod-ed25519-20260523 and externally timestamped Sep 11, 2026, 4:00 PM.
sha256:00fe91fca9a674...d5af0b0adcPreseen researches your questions and answers them with calibrated, source-backed forecasts.