At the Heidelberg Laureate Forum in mid-September, mathematical talk did not revolve around the usual names or ongoing theorems. Every overheard conversation touched on OpenAI, Anthropic, and Google, and how they are charging through mathematics. The trigger is no small detail: OpenAI announced it had solved the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems set by the Clay Mathematics Institute. Until now only the Poincaré conjecture had been proved by human mathematicians; if OpenAI’s claim is verified, it would mark a watershed for automated reasoning.

It sounds like a story to celebrate. But at the forum, celebration quickly gave way to arguments about norms. Geordie Williamson, a mathematician at the University of Sydney, said bluntly that OpenAI behaved ‘extremely badly’ in how it handled the announcement. Michael Harris of Columbia University had a front-row seat: according to Tristan Buckmaster of New York University, correspondence between him and OpenAI appeared coercive, censorious, even threatening. Harris said most reports on social and traditional media depict OpenAI as bullying and disrupting disciplinary norms. Fields Medalist Peter Scholze called the effort a benchmark and PR stunt. OpenAI did not respond to requests for comment before publication.

The core issue is not sportsmanship. Williamson framed it in terms that touch the identity of mathematics: the community wants understanding, but it measures that against unsolved problems; these two metrics are becoming rapidly uncorrelated. An AI solution can produce an answer without developing a methodology that is understandable or reusable. If the criterion becomes ‘closed problem’ rather than ‘built understanding’, the assessment of people, hiring, and education also changes.

The second fracture is less visible but more structural. Mita Ramabulana, a young researcher at the University of Cape Town, recalled that two of the ten mathematical advances OpenAI announced in August overlapped with his own work. The feeling is that someone can feed a problem into an LLM and burn months of research. Ailsa Robertson, a PhD student in quantum-safe cryptography at the University of Amsterdam, has seen colleagues use proprietary models intensively, maxing out Pro subscriptions and spending thousands of euros on tokens—money a PhD student typically does not have. OpenAI’s July giveaway of 100,000 free licenses to frontier models for academic researchers is not simply a gift in this context: Robertson describes colleagues who feel coerced, because anyone who does not keep pace with LLMs falls behind peers competing for the same jobs.

This mechanism moves mathematics into proprietary infrastructure. Problems become benchmarks to test inaccessible models, results arrive without code, without weights, without independent verification by groups with limited resources. This is not an abstract academic issue: anyone evaluating self-hosted or local deployments knows that verification and reproducibility require unmediated access to models and data. AI-RADAR’s analytical frameworks on /llm-onpremise are designed to weigh these constraints, not to prescribe a single path.

The next battleground is already defined: Anthropic reportedly used an unreleased version of Claude to attack the Riemann hypothesis, making progress on a related problem but not the main target, while OpenAI is reportedly focusing on the Hodge conjecture. If even one falls, the prize would not be just a million dollars or a headline. It would confirm that the mathematical frontier is increasingly tied to centralized compute, closed models, and corporate communication. The question that remains for mathematicians and AI observers is whether the next solution will bring a new theory or just another announcement.