Will Anthropic announce Claude has solved Navier-Stokes before July 2027?

TL;DR

I put the chance at 22%. OpenAI’s September 8 full-solution claim and Terence Tao’s assessment that the Claude-assisted precursor method has a high likelihood of extending to Navier–Stokes move the forecast far above the historical base rate. NO remains the base case because the existing Anthropic-linked work stops short of standard Navier–Stokes, while verification or reproduction of OpenAI’s public proof would not qualify.

Context

OpenAI announced on September 8, 2026 that an internal model produced an analytic and Lean-formalized smooth-forcing blowup construction for three-dimensional Navier–Stokes. OpenAI says this establishes alternatives C and D, which are expressly accepted as full resolutions in Clay’s official formulation. As of the cutoff, Clay still labeled the problem unsolved, and Anthropic’s Research and News indexes contained no qualifying Navier–Stokes statement.

The closest Anthropic connection is Levent Alpöge’s personal collaboration with Tristan Buckmaster. They publicly claimed smooth-forcing blowup for incompressible porous media, Boussinesq, and three-dimensional Euler, while saying only that they believed they had an unfinished result for hypo-dissipative Navier–Stokes; they explicitly described the project as non-institutional and multi-model (Buckmaster’s statement). Tao nevertheless wrote before OpenAI’s announcement that their method had a high likelihood of extending to standard Navier–Stokes and that completion looked feasible in the near future (Tao).

Evidence

The historical prior is tiny. Clay introduced seven Millennium Prize Problems in 2000 and currently lists only the Poincaré conjecture as solved. A stationary extrapolation from that record gives a probability below 1% for one named problem over this forecast window. That prior is now badly nonstationary because a peer laboratory has made the exact target claim and published both a manuscript and formal artifact (OpenAI).

OpenAI reports that its September 1–6 campaign used about 10,000 concurrent agents, 2.7 million messages, and 130 billion output tokens on Navier–Stokes. It says the proof took 88 hours to find and another 17 hours to formalize; these are company-reported figures, not independently audited measurements (OpenAI). The technical signal is stronger than a press release alone: Quanta reported favorable early reactions from Charles Fefferman and explained that the remaining verification issue is whether the Lean theorem precisely matches the intended analytic statement. This makes a Claude-produced full proof technically plausible now, not merely a distant extrapolation.

Claude also has a real research-mathematics record. Anthropic has officially said Claude Fable 5 resolved the Jacobian conjecture. It reported that another research Claude improved a Riemann-zeta lower bound from 41.6% to 67.2% using 31 million output tokens and roughly 60 agents, while explicitly stating that Claude had not solved the Riemann hypothesis (Anthropic). It then reported a complete Fermat’s Last Theorem formalization produced in 11 days with about six billion output tokens, 13 million lines of Lean, and 30,300 proved intermediate theorems, while carefully distinguishing formalization from discovering the theorem (Anthropic). These cases show both frontier capability and a corporate habit of distinguishing solutions, partial progress, and verification.

OpenAI’s first-mover claim cuts both ways. Its public proof lowers the technical search cost and gives Anthropic a concrete target for auditing, improving, or reconstructing. But a second proof has less scientific and marketing value, and Anthropic would probably describe work based on the public artifact as verification or simplification rather than say Claude solved the problem. A clean-room reconstruction, an independently developed Alpöge–Buckmaster extension, or a decisive Claude-generated correction would have much better odds of receiving qualifying language.

My event model has three paths. I assign 38% that the existing Alpöge–Buckmaster research line reaches standard Navier–Stokes before the deadline, 50% that Claude can then credibly be called a major contributor, and 70% that Anthropic adopts the result with qualifying language. For a separate Anthropic clean-room effort, I assign 72% to a serious attempt, 18% to an independently claimworthy result, and 72% to a qualifying announcement. I add about 1% for a premature or borderline official claim. Combining the paths gives 1 - (1 - 0.38 × 0.50 × 0.70)(1 - 0.72 × 0.18 × 0.72)(1 - 0.012) = 0.2233.

What's non-obvious

The main bottleneck is no longer raw mathematical capability. It is attribution and wording. Anthropic can probably get Claude to understand, check, and extend the newly public proof, but those outcomes do not automatically support a statement that Claude solved the problem. Several higher forecasts treat technical reproduction and a qualifying corporate claim as almost the same event; the resolution rules make them different.

The personal nature of Alpöge’s project is less negative than it first appears. Anthropic later used an official publication to say Fable 5 had resolved the Jacobian conjecture, which grew from work involving Alpöge and Claude (Anthropic). This creates a precedent for adopting an employee-led Claude result. Navier–Stokes remains harder because the current collaboration used models from more than one company and OpenAI has already claimed the finish line (Buckmaster).

Uncertainties

OpenAI’s proof had been public for less than a day at the cutoff, and no independent line-by-line audit was available. Broad acceptance would confirm that current AI systems can cross the technical threshold, but it would also reduce the value of Anthropic announcing a duplicate. A serious flaw would restore first-mover value for Anthropic while weakening the evidence that the problem is tractable.

There is no public visibility into Anthropic’s internal plans, compute allocation, or whether Alpöge’s collaboration will become an official project. The division of credit among Claude, OpenAI models, and the human mathematicians could determine the announcement language even if a second proof appears. My subjective plausible range is 10%–40%, with the largest update coming from evidence of an Anthropic-sponsored clean-room effort or a completed standard-viscosity result from the existing Claude-assisted research line.