Overview
- OpenAI said on September 8 that an unreleased internal model produced a proof for the Navier‑Stokes existence and smoothness problem after about 10,000 AI agents worked for roughly 88 hours and GPT‑6 Astra performed a Lean formal verification in about 17 hours.
- Two researchers who had recent related work, Tristan Buckmaster and Levent Alpöge, questioned the timing and possible use of material stored in OpenAI systems, and OpenAI said it cannot rule out that anonymized product data helped improve models.
- Demand for GPT‑6 Astra has surged since its limited rollout on September 3, and OpenAI product lead Tibo Sottiaux warned that the company may temporarily pause new Pro subscriptions to protect existing users.
- Reported compute costs for the proof effort are estimated at $15 million to $22.5 million, and OpenAI told reporters it will not pursue the Clay Mathematics Institute’s $1 million prize for the result.
- The announcement has spurred calls for outside peer review and reproducibility, prompted at least one high‑profile resignation and raised broader concerns among mathematicians and safety researchers about data governance and the role of AI in frontier discovery.