We already covered OpenAI's September 8 announcement: 10,000 agents produced a proof concerning singularities in the Navier-Stokes equations. What we didn't report at the time, because it wasn't confirmed yet, is the uncomfortable part that surfaced afterward: a Fortune investigation revealed a priority dispute with two outside researchers, intimidation allegations, and a direct public rebuke from Terence Tao — one of the most respected living mathematicians — over how AI companies are handling this kind of announcement.
One detail that got lost in the initial coverage: the reasoning model that produced the mathematical proof wasn't GPT-6 Astra — it was an internal OpenAI model reported as "GPT-5.6 Sol," not yet available to the public. Astra's confirmed role in this project was different and more limited: formalizing and auditing the proof in Lean (the formal proof-verification language), not the original mathematical reasoning. It's a distinction that matters — "a model solved the problem" is not the same claim as "a model verified that another model's solution was logically valid."
To date, there is no public, conclusive evidence that confirms or rules out that allegation beyond Bubeck's denial — we mention it because it's a legitimate part of covering this story, not because we treat it as proven.
The most serious part of Fortune's reporting concerns allegations of direct pressure: according to the investigation, Bubeck allegedly demanded that Alpöge's name be removed from any claim related to the problem's prize, citing his affiliation with Anthropic, and reportedly went as far as asking him: "Why would you ruin your career?" These are serious allegations, fitting for a dispute between two of the world's largest AI labs, and still without a clear public resolution as of this writing.
Terence Tao, a Fields Medal winner and one of the most-cited voices in pure mathematics, publicly criticized the broader pattern of AI companies "exploiting" the Millennium Prize Problems for marketing purposes without necessarily producing new mathematical understanding. His comparison was blunt: using these announcements this way is like "using bulldozers to loot an archaeological site" — you can extract something of value quickly, but you destroy the context and process that would have given it real long-term scientific value.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies in Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel