Navier–Stokes finite-time blow-up — Millennium Prize problem
Claimed proof of finite-time blow-up for the forced Navier–Stokes variants (C)/(D) of the Clay formulation, produced in ~88 hours with a Lean formalization.
Updated weekly · Edition: November 15, 2026 — What AI has actually claimed, what it has partly cracked, and what has been checked line-by-line. The distinction matters more than the headline count.
AI produced a complete claimed solution. Not yet formally verified or independently confirmed.
Claimed proof of finite-time blow-up for the forced Navier–Stokes variants (C)/(D) of the Clay formulation, produced in ~88 hours with a Lean formalization.
Announced as solving 5 of 6 IMO 2025 problems in contest time for a gold-level score, graded by three former IMO medallists.
OpenAI stated an internal model resolved 100+ open problems since late August 2026 but published no problem list and no proofs — unverifiable as stated, resting entirely on the lab's announcement.
Complete claimed solutions from the same release, not yet formally verified or independently confirmed. Source: github.com/openai/math
Special cases, improved bounds, weakened versions, or AI contributions that humans completed.
Found a configuration of 593 spheres kissing a central sphere, improving the long-standing 592 lower bound. The exact number is still open (upper bound 868). Autonomous agents later raised the bound further, to 604.
Improved the 49-multiplication record that had stood since Strassen (1969). The general exponent problem remains open.
A Markov-chain method suggested by the model unlocked all three problems, but the proofs were constructed and verified by the human co-authors — AI-assisted, not autonomous.
Produced a Lean proof in ~6 hours, but Tao and Bloom noted Erdős had omitted a key hypothesis from the statement, so the solved version is substantially weaker than the intended problem.
On ten unpublished research-level problems, DeepMind claims 6 of 10 solved autonomously; the organizers judged only 2 of 10 definitively solved, and one duplicated an existing proof sketch. Lab and organizer assessments conflict — contested partial progress.
Jiesong Zhang's paper proves the codimension-one case of Étienne Ghys's 1995 conjecture on holomorphic Anosov diffeomorphisms, which includes the full complex-dimension-three case; the general conjecture in higher codimensions remains open. Zhang disclosed in the paper's October 13 revision (arXiv 2609.37695, first posted September 29) that he developed ideas and arguments over multiple rounds with Shanghai AI Lab's internal Intern-S2 model and independently checked the mathematics himself. A separate Peking University team (Dong Bin and Liu Ruochuan, using their SHEAF framework) formalized the paper's central results in Lean — over 550,000 lines — with human verification. It sits in Partially Solved because only a special case is proved and the AI's role was assistive, not autonomous; headlines claiming the full 31-year-old conjecture was solved overstate the result.
The paper proves that for every k ≥ 3, any family of subsets of [n] with no k-sunflower has size at most ckn for some ck < 2, settling the Erdős–Szemerédi sunflower conjecture from the 1970s. In the authors' own statement on the use of AI, the proof of the key step, Theorem 2, was found by ChatGPT 6 Astra Pro after a short discussion generalizing the k=3 argument; the authors rewrote it in polynomial-method language and verified all results themselves. Honest note: the authors also state the proof has been verified in Lean 4.35, but that verification is not yet publicly available (to be released on the Palomar repository), so it is not counted here as proven and confirmed.
The 1995 conjecture asks whether every isotropic log-concave probability measure has a Cheeger constant bounded below by a universal positive constant; the paper proves ψn ≤ C for a universal constant, removing all dimension dependence. It was produced by a human-directed agentic AI process in which AI agents explored directions and humans directed, pruned, and completed the proof. Honest notes: the paper is not peer-reviewed; two parallel proofs appeared the same day (Bizeul, Klartag and Lehec, who disclose most proofs and ideas were found by ChatGPT; and Balasubramanian and Kasiviswanathan on the dimension-free Poincaré constant); and a Lean formalization of a dimension-free Poincaré bound (KLS500) exists, but it is not a full published Lean check of the Song–Zhang paper.
Formally machine-verified (Lean / Coq) or independently confirmed by mathematicians or peer review.
OpenAI’s Oct 26 claim that its models produced rigorous existence proofs for interacting four-dimensional quantum field theories — including pure Yang–Mills theory — is now verified: a committee led by Nobel laureate David Gross released the Lean formalization on Nov 10, 2026, and Scientific American confirms the proof checks out. The existence half of the Yang–Mills Millennium Prize problem is now proven by AI; the mass-gap half is separate and is not claimed here.
Complete proof produced with 64 subagents in under an hour; kernel-checked in a public Lean repository; two senior graph theorists (Geelen, Oum) independently published expositions within a week. Not yet journal peer-reviewed.
Constructed point sets with polynomially more unit-distance pairs than the square grids long believed optimal; nine external mathematicians independently verified it and published a companion paper the same day.
GPT-5.2 Pro produced the proofs, Harmonic's Aristotle formalized them in Lean; #728 is regarded as the first Erdős problem fully resolved autonomously by AI, and Terence Tao accepted the #397 proof as an original result (while cautioning these are “lowest-hanging fruit”).
Solved 9 of 353 formalized open Erdős problems (two unsolved for 56 years) and proved 44 OEIS conjectures, plus an algebraic-geometry problem and a minimax bound — all Lean-checked and published through peer review in Science.
Arguments for ten problems open a decade or more (also non-sofic groups, Ehrhart's Volume Conjecture, sphere-packing and coding bounds, others), all formalized as Lean 4 certificates with zero “sorry” placeholders, in a 249-page manuscript. Journal peer review still pending.
Settled the problem in the negative with essentially no human mathematical input, backed by a ~19,000-line Lean formalization that builds cleanly and passes a statement-matching check; human experts verified eight Rethlas results including this one.
Proved the conjecture unconditionally via a Jacobi-symbol reformulation.
Deep Think scored 35/42, a gold-medal performance officially certified by the IMO, solving 5 of 6 problems; Aristotle independently produced Lean-verified proofs for 5 of 6.
On Oct 26, 2026 Anthropic posted “A 2^o(n) Time Algorithm for 3-SAT”, credited to “Claude (Anthropic)”: an internal research Claude model found the core deterministic 3-SAT algorithm on Oct 16 (the same day OpenAI reports) using 14 million output tokens with no human input; it takes a different route (tensor-network contraction, an 1840s eight-square identity from the octonions); the Lean 4 formalization was written by AI models; the paper cites OpenAI's preprint as independent and concurrent work. No practical speedup (it beats brute force only at enormous problem sizes) and it does not settle P vs NP. No human has fully confirmed it — Anthropic's claim.
On Oct 26, 2026 the startup Axiom announced the first complete computer-checked proof of the Classification of Finite Simple Groups (the “Enormous Theorem”) in Lean: about 43.1 million lines, sorry-free, using only Lean's standard axioms, produced in 24 days by swarms of AI agents directed by engineers and mathematicians. The formalization exposed 34 noteworthy deviations in the published literature, including a counterexample (PSL₂(49)) to a step in Walter's 1986 Unbalanced Group Theorem proof; Axiom's AxiomProver system proved a new group-theory result (“Preservation of simple tops”, Theorem 2.3, with Colin Defant and Ken Ono) to patch it. The code is released for independent reproduction but has not yet been independently rebuilt. Caveat: this is a formalization of the long-established human proof (some 15,000 pages by about 100 mathematicians), not a newly discovered solution — Axiom's claim.
On Nov 6, 2026 PSI announced its internal research system, running instances of a non-public Anthropic model, produced a complete proof of the Aaronson–Ambainis conjecture on quantum speedups by the third day of a run, then improved the bound twice; the Lean 4 formalization (four main theorems, no placeholders, no added axioms, re-checked with two independent Lean kernel implementations) and the paper are public on GitHub. The conjecture, posed by Scott Aaronson and Andris Ambainis, says every bounded low-degree polynomial on the Boolean cube with non-trivial variance has an influential variable — ruling out exponential quantum speedups on random inputs. Not peer-reviewed; experts must still confirm the formal statement matches the conjecture as intended. Co-proposer Scott Aaronson blogged Nov 8, 2026 that he believes the proof is correct (Lean certificate plus expert review, including a Stanford exposition by Blanc, Docter, Strassle and Tan), while noting he has not yet checked it himself.
Source: the release repo, github.com/openai/math; ‡ = in the repo's stricter formalized-main-result catalogue — 176 papers with a formalized main result, spanning 139 families; a “Lean” link points to the family's formalization notes in the repo; includes formalized bound results such as the unit-distance power-saving bound, where the underlying problem can remain open even though the stated result is machine-checked.
Why this tracker splits claims from checks — four cautions that apply across the whole page.
First edition · October 13, 2026 · Sources linked on every entry — check the source before citing a result as settled.
Next weekly update will re-check verification status and move entries between categories as proofs are confirmed or withdrawn.