On the Navier–Stokes Millennium Prize Problem

OpenAI says an internal model — meaningfully more capable than GPT-6 Astra and still mid-training — resolved one of math's seven Millennium Prize Problems: whether a smooth, finite-energy 3D fluid can develop a singularity (unbounded speed) in finite time under the Navier–Stokes equations, open since the 1930s. The claimed proof came from roughly 10,000 coordinating agents working ~88 hours (2.7M messages, ~130B output tokens), with a Lean-formalized, machine-checked version following over another 17 hours; OpenAI is not claiming the Clay Institute's prize money, framing this as a capability progress report.

The release lands amid a live priority dispute: NYU mathematician Tristan Buckmaster — who was independently working the related Euler-equations problem with Anthropic researcher Levent Alpöge — says OpenAI converged on his unpublished approach around the same time he informed the company of his own progress, and that OpenAI pressured him over co-author credit; he also worries his heavy use of Codex could have leaked signal into OpenAI's training data. OpenAI says it never saw the pair's unpublished work before public release and that the two proofs differ substantially, while conceding it can't fully rule out that de-identified usage data helped. Terence Tao's public reaction frames the deeper worry: that AI labs racing to "flatten" a problem the moment they hear a rumor someone is close could push researchers to stop sharing promising directions at all.

Introducing ChatGPT Images 2.5

OpenAI shipped ChatGPT Images 2.5 plus two new API models — GPT-Image-2.5 Flare (default, 50% lower latency than GPT-Image-2) and GPT-Image-2.5 Sunburst (higher-precision, slower) — with sharper detail, better reference-photo fidelity, and edits that stick across multi-turn conversations instead of drifting. New product features: Sketch (draw a rough reference directly in ChatGPT), format templates, on-image comments, and prompt-sharing. Rolling out today to all ChatGPT/ChatGPT Work/Codex tiers; the two API models are live now with published pricing.

AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome

DeepMind released a free, public atlas precomputing its AlphaGenome model's predicted molecular effect for all ~9 billion possible single-letter DNA variants in the human genome — a 1-petabyte dataset, over 30x the size of the AlphaFold Database. Alongside the raw predictions, DeepMind is releasing an "AVI" score that folds in AlphaMissense (its protein-impact model) to give researchers one number to rank any variant's likely impact, covering both the 2% of the genome that codes for proteins and the 98% that doesn't. Available today via a web portal, the AlphaGenome API, and as a skill in Google Antigravity; DeepMind says external collaborators have already used it to find and experimentally verify rare-disease variants.