OpenAI Claims a Millennium Prize Proof While India’s Institutes Are Shut Out of Free Frontier-Model Access

The news

Bengaluru. On September 8, OpenAI announced that an unreleased internal model had solved the Navier–Stokes existence and smoothness problem, one of six unsolved Millennium Prize Problems, each carrying a million-dollar bounty, The Hindu’s science page reports. According to the report, OpenAI, spurred by rumours that its rival Anthropic was closing in, deployed 10,000 autonomous AI agents and millions of dollars’ worth of computing power and found a solution in 88 hours. The claim is already contested. On September 7, New York University mathematician Tristan Buckmaster revealed that he and Levent Alpöge had worked for about a year on a simplified version of the problem using similar methods, assisted by both companies’ models; OpenAI denied specifically accessing their data but said it “cannot rule out” that de-identified data from their usage helped improve its models. The run of claims began earlier. On May 20, OpenAI said its internal reasoning model had solved the planar unit distance problem posed by the Hungarian mathematician Paul Erdős in 1946. On July 20, Mr. Alpöge, who works at Anthropic, posted on X that its model Claude Fable 5 had produced a counterexample — a single case that proves a general claim false — to the Jacobian conjecture, an 87-year-old open problem in algebraic geometry. On August 1, OpenAI said its newest model, ‘Astra’, had solved ten more open problems, three from Erdős’s list. Mathematicians are divided. Bristol’s Andrew Booker says Claude Code wrote 30,000 lines of code and about 200 pages of mathematics in a week on a problem he had been stuck on for about ten years, and MIT’s Andrew Sutherland calls it “an exciting time”; yet Dr. Sutherland warns that AI proofs gloss over the key intuitive step, and Dr. Booker that, left unchecked, they can become “completely incomprehensible”. Columbia’s Michael Harris and UC San Diego’s Karthik Ganapathy, who helped draft the June 2026 Leiden Declaration on Artificial Intelligence and Mathematics, fault companies for announcing results without their methods. On July 29, OpenAI offered 1,00,000 scientists, mathematicians and engineers free access to its frontier models but excluded Indian, Russian and Chinese institutes, the report says, most likely because of geopolitical tensions. It warns that India, with no frontier model of its own, may struggle to keep pace in theoretical science. The syllabus link is GS3 on developments in science and technology, computers and the indigenisation of technology.

Static syllabus linkage

1. The Clay Institute priced seven problems at $1 million each in 2000, and only one has been solved. The Clay Mathematics Institute, a private foundation in the United States, announced the seven Millennium Prize Problems in Paris in 2000 and offered US$1 million for the solution of each. They are the Birch and Swinnerton-Dyer conjecture, the Hodge conjecture, Navier–Stokes existence and smoothness, P versus NP, the Poincaré conjecture, the Riemann hypothesis and Yang–Mills existence and mass gap. Only the Poincaré conjecture has been solved, by the Russian mathematician Grigori Perelman in 2002–03; he declined the Fields Medal in 2006 and the Millennium Prize awarded in 2010, which is why the report speaks of six remaining problems. The Institute’s rules require publication in a qualifying outlet, a wait of at least two years and general acceptance by mathematicians before any award, so a company’s announcement is a claim, not a prize.
2. The three problems in the news come from three different branches of mathematics. The Navier–Stokes equations, named after Claude-Louis Navier and George Gabriel Stokes, describe how viscous fluids such as air and water move, and underpin weather forecasting and aircraft design; the prize asks whether their three-dimensional solutions stay smooth for all time or can ‘blow up’. The Jacobian conjecture, posed by the German mathematician Ott-Heinrich Keller in 1939, states that a polynomial map whose Jacobian determinant is a non-zero constant must have a polynomial inverse, so one counterexample disproves it. The planar unit distance problem, posed by Paul Erdős (1913–1996) in 1946, asks for the largest number of pairs of points exactly one unit apart among n points in a plane. Erdős left hundreds of such problems, which is why they have become a yardstick for AI systems.
3. A proof assistant certifies that a proof is correct, not that anyone understands it. A proof assistant such as Lean, created by Leonardo de Moura at Microsoft Research from 2013, requires a proof to be written in a precise formal language so that a small trusted program can check every step; the report says current models can produce proofs and formalise them in this way. Computer-assisted proof is not new: Kenneth Appel and Wolfgang Haken’s 1976 proof of the Four Colour Theorem relied on a computer to check cases no human could, and Georges Gonthier verified it in the Coq proof assistant in 2005. The lesson of that history is that machines can settle whether a result is true, while the understanding of why it is true still has to be built by people.
4. India has built shared compute under the IndiaAI Mission, but not a frontier model. A frontier model is a general-purpose AI model at the leading edge of capability, trained with very large computing power and data; the term was used in the Bletchley Declaration of November 2023, which India signed. The Union Cabinet approved the IndiaAI Mission on March 7, 2024 with an outlay of ₹10,371.92 crore, implemented by the IndiaAI Independent Business Division under the Digital India Corporation, across seven pillars that include compute capacity, an innovation centre for indigenous models and safe and trusted AI, per PIB. It targeted 10,000 or more GPUs, the chips used to train and run AI; a PIB release of March 25, 2026 says more than 38,000 have been onboarded for start-ups and academia at affordable rates. Shared compute lowers the cost of using models, but training one at the frontier needs far larger, concentrated investment.

Why UPSC loves this

1. The GS3 syllabus asks about developments in science and technology and their effects. GS3 lists science and technology developments and their effects in everyday life, awareness in IT and computers, and the indigenisation of technology. Prelims 2020 asked what AI could effectively do at its present stage of development, and Mains 2023 asked how AI helps clinical diagnosis and whether it threatens privacy. This story moves the question to how AI changes the way knowledge is produced and certified.
2. Access to frontier AI is a running thread in this magazine. The 22 September card covered calls for India to secure early access to frontier models for safety testing, and the 23 September card the Xi–Trump AI dialogue. Today adds the science angle: a company’s access list now decides who can compete in basic research, which serves GS2 answers on technology diplomacy and GS3 answers on indigenisation.

Analysis

1. Until it is verified and understood, the Navier–Stokes result is a corporate claim, not a theorem. The report says OpenAI ‘announced’ a solution; it does not say the proof has been formally verified, refereed or published. The Clay rules demand publication, a two-year wait and general acceptance precisely because famous problems attract flawed proofs. A formalisation in Lean would settle correctness, but the mathematicians quoted say AI proofs skip the hardest steps and drift into invented terminology, so human checking could take years. The counter-view is real: experts who checked the Erdős unit distance solution judged it worthy of a top journal, so these systems are not producing noise. The sound posture is neither dismissal nor celebration but a firm line between ‘announced’, ‘verified’ and ‘accepted’.
2. Secrecy about prompts and methods breaks the norm that makes science self-correcting. Science corrects itself because others can reproduce the method, not merely admire the answer. Harvard’s Nina Zubrilina asks for the prompts and objects to ‘intentional mysticism’; without them nobody can tell how much the model did and how much the company’s own expert mathematicians did. Michael Harris’s reading, that companies chase mathematics to signal progress towards artificial general intelligence and to please investors, explains the incentive: the headline is the product. The counter-view is that companies have spent compute no university could afford and that published answers still invite scrutiny; the Leiden Declaration is the community setting terms for the tools rather than refusing them.
3. The Buckmaster episode warns that researchers’ unpublished ideas can become training data. Tristan Buckmaster and Levent Alpöge worked for a year on a simplified Navier–Stokes problem using both companies’ tools, and OpenAI ‘cannot rule out’ that their de-identified usage data improved its models. A researcher using a commercial model on unpublished work may thus be feeding a system its owner can point at the same problem. De-identification protects the person, not the idea, and a research idea is not ‘personal data’ under India’s Digital Personal Data Protection Act, 2023. Universities and funders therefore need their own rules on AI use for unpublished work and on contractual opt-outs from model training. The counter-view is that usage data improves tools for everyone, but priority and credit are the currency of academic careers and deserve protection.
4. Excluding Indian institutes shows that access, not only ownership, is now an instrument of power. The report groups Indian institutes with Russian and Chinese ones and attributes the July 29 exclusion ‘most likely’ to geopolitical tensions; no formal reason is given. The pattern is familiar: a U.S. AI diffusion rule of January 2025 placed India in a middle tier of countries facing caps on advanced chips, before it was rescinded in May 2025. A company’s access list is not an export control in law, but it has the same effect on an Indian doctoral student. The Economic Times reports that French President Emmanuel Macron urged countries at the UN General Assembly to pool investment in an open-source frontier model independent of the great powers, one route for a middle power. The counter-view is that paid and open models remain available, so the exclusion is a cost, not a wall; but where, in Dr. Booker’s experience, two weeks of AI help beat five years of effort, a cost decides who finishes first.
5. India’s edge in theoretical science will erode unless model access is treated as research infrastructure. The report notes that India once enjoyed an edge in the theoretical sciences, an advantage built on talent and modest inputs rather than expensive laboratories. If a model is a ‘grad student army’, a mathematician’s output now depends on access to compute and models as much as on training. The IndiaAI Mission’s GPUs are open to academia, but theory departments rarely compete for such resources, and the Anusandhan National Research Foundation, set up under a 2023 Act, could earmark model access for basic research. Training must also shift from pure problem-solving towards exposition, formalisation and judgement, as Dr. Sutherland suggests. The counter-view is his reassurance that mathematics will never be ‘solved’; yet the report’s note that 2026 Fields medallist Jacob Tsimerman stopped taking graduate students shows how seriously the field takes the risk.

Administrator's brainstorm