AI Math Breakthroughs 2026: What’s Proven So Far

Quick answer: In 2026, AI systems moved from olympiad problems to contributing to famous research questions. The biggest results: an AI-found disproof of Erdős’s unit distance conjecture (May); a proof of the cycle double cover conjecture (July); a counterexample to the Jacobian conjecture in three variables (July); a Navier–Stokes blow-up result with an external force (September); a Lean formalization of Fermat’s Last Theorem (September); and a Lean proof of the Riemann Hypothesis developed with Claude, certified by the Lean/Mathlib community on November 21. Several other claims, including refutations of the Exponential Time Hypothesis, were still under review in late November.

Evidence last checked: November 22, 2026.

For most of history, a famous math problem fell perhaps once a decade. In 2026, headlines about AI and famous problems arrived almost monthly. Some results are solid, some are still being checked, and all of them are easy to misread. This page collects the biggest AI math breakthroughs of 2026, explains each in plain words and gives its status. We chose them by mathematical significance, novelty, strength of the evidence and importance for how mathematics is done.

2026 Timeline

Public dateResult and scopeHuman and AI contributionAvailable checks
Jan 2026Erdős Problem 728 solvedKevin Barreto with GPT-5.2 Pro; Lean proof by Harmonic’s AristotleLean-checked
Feb 11Google DeepMind’s Aletheia agent: autonomous solutions to four of 700 Erdős problems it evaluatedGemini Deep Think research agentPublished by DeepMind with papers
May 20Erdős’s 1946 unit distance conjecture disprovedOpenAI model; companion paper by nine mathematiciansHuman-verified companion argument
JulyCycle double cover conjecture (1970s) provedOpenAI’s GPT-5.6 Sol; human exposition by Jim Geelen and othersExpository paper; expert commentary
JulyJacobian conjecture disproved in three variablesCounterexample credited to Anthropic’s Claude, announced by Levent AlpögeIndependent verification record
Aug 10More zeta zeros proved on the critical line: lower bound raised from 41.6% to 67.2%Claude research model; checked by Anthropic mathematicians, examined by outside expertsPaper plus Lean formalization
Sept 4Fermat’s Last Theorem fully formalized in Lean (about 13 million lines in 11 days)Claude, set up by Anthropic researchers; reviewed by Kevin BuzzardLean-checked; a checking milestone, not a new proof
Sept 8Finite-time blow-up for 3D Navier–Stokes with a smooth external forceOpenAI agentsLean formalization; Clay said the problem appears settled
OctQuasi-Riemann hypothesis provedOpenAIFormal materials released; expert study ongoing
Oct 6OpenAI releases about 720 research manuscripts at different stages of verificationOpenAI systemsMixed; many unreviewed
Oct 26Claimed refutations of the Exponential Time HypothesisOpenAI and Anthropic, separatelyLean code compiles; community examination ongoing
Nov 20Lean proof of the Riemann HypothesisLevent Alpöge, developed with ClaudeStatement and proof certified by the Lean/Mathlib community on Nov 21
Nov 20Navier–Stokes blow-up with no external forceAlpöge and Tristan Buckmaster, with heavy AI useAnnounced; public verification package incomplete

Did AI Prove the Riemann Hypothesis?

The Riemann Hypothesis (1859) says all nontrivial zeros of the zeta function lie on the critical line, and it controls how regularly primes are spread out. In 2026 AI moved toward it in three steps:

DateResultWhat it established
Aug 10Critical-line proportionAt least 67.2% of zeros lie on the line (up from 41.6%)
OctQuasi-Riemann hypothesisAll zeros stay in the strip with real part between 1/8 and 7/8
Nov 20Full Riemann Hypothesis in LeanAll nontrivial zeros on the line, as a machine-checked proof

On November 20, Alpöge released a Lean proof of about 17 million lines developed with Anthropic’s Claude. The next day the Lean and Mathlib communities certified that the formal statement matches the standard one and that the proof checks; MIT’s Andrew Sutherland logged several successful independent checks, with a few extra runs still in progress. A human-readable explanation has not yet appeared. The certification covers the classical RH and a version for Dirichlet L-functions, not every generalized form. The Clay Institute still lists RH as unsolved: its prize rules require publication, a two-year wait and general acceptance, a separate process from formal verification.

Did AI Solve Navier–Stokes?

The Navier–Stokes Millennium Problem asks whether smooth 3D fluid flows can develop infinite speed in finite time. On September 8, OpenAI announced an AI-found proof of blow-up when a smooth external force pushes the fluid, one of the cases Clay’s official problem statement allows; Clay said the problem appears to have been settled. On November 20, Alpöge and Buckmaster announced blow-up with no force at all, found with heavy AI use according to Terence Tao, but some numerical verification materials were not yet public.

Other Major Discoveries

  • Unit distance conjecture (May). An OpenAI model found a construction beating Erdős’s conjectured growth rate; Tim Gowers called it unlike any previous AI proof. See Erdős problems solved by AI.
  • Cycle double cover conjecture (July). A graph-theory conjecture from the 1970s about covering every edge of a graph exactly twice by cycles. OpenAI’s GPT-5.6 Sol produced a proof, and mathematicians including Jim Geelen wrote an exposition making it readable.
  • Jacobian conjecture (July). A three-variable polynomial map with constant nonzero Jacobian determinant that sends different inputs to the same output, which disproves the conjecture in three or more variables. See Jacobian conjecture disproved.
  • Erdős problems at scale. Google DeepMind’s Aletheia agent (February) and later systems settled several listed problems, while many other claims proved to be rediscoveries.

Proof-Checking Milestones

Not every breakthrough is a new theorem. In September, Claude formalized the entire proof of Fermat’s Last Theorem in Lean in 11 days, about 13 million lines and 30,000 theorems. Kevin Buzzard, who had led a multi-year human effort to do the same, reviewed it. This does not discover anything new about Fermat, but it shows AI can turn enormous human proofs into machine-checked ones, which is how the Riemann Hypothesis proof could be certified so quickly.

What Would an ETH Refutation Change?

Computer scientists widely assumed that the 3-SAT problem cannot be solved much faster than brute force (the Exponential Time Hypothesis). On October 26, OpenAI and Anthropic each announced AI-found, slightly subexponential algorithms that would refute it. Russell Impagliazzo, Ramamohan Paturi and Ryan Williams organized community examination but said their involvement is not certification. Even if confirmed, it would not settle P vs NP.

Has AI Solved a Millennium Prize Problem?

Problem2026 AI-assisted resultClay status (Nov 22)Prize
Riemann HypothesisLean proof certified by the Lean/Mathlib communityListed as unsolvedNone announced
Navier–StokesForced blow-up (Sept); unforced claim (Nov)Said to appear settled; listed as activeNone announced; OpenAI said it will not claim it
P vs NPNone (ETH claims are a different question)UnsolvedNone

What These Results Have in Common

  • Explicit checking. Every result needed checking, but by different methods: Lean for some, human companion papers for others, and some packages are still incomplete.
  • Human roles changed, not vanished. People chose the problems, checked that statements matched, and wrote expositions.
  • Understanding lags proof. In the Navier–Stokes case, NPR’s September headline put it bluntly: “Humanity learned nothing (so far).”

How to Read the Next AI Math Headline

  1. What exactly was proved: the original problem or a variant?
  2. Is there a formal proof, and has someone checked that its statement matches the real problem?
  3. Have independent experts reviewed it, or only the company that announced it?
  4. Is it new, or a rediscovery of an older result?

For the basics of what AI can and cannot do, see Can AI Do Math?

Common Misconceptions

“AI solved P vs NP.” No. The October claims concern the Exponential Time Hypothesis, a different, stronger assumption.

“The Riemann Hypothesis prize was awarded.” No. Formal verification is done, but Clay’s prize process has separate rules and a waiting period.

“Formalizing Fermat’s Last Theorem means AI discovered it.” No. Andrew Wiles proved it in 1994; AI produced a complete machine-checked version in 2026.

Watch: 2026 AI Math Breakthroughs in Video

If you learn better by watching, these videos are worth your time.

Did AI Solve One of Math’s Hardest Problems? (StarTalk). A popular discussion of the 2026 AI math claims and what they mean.
What Really Happened When AI Solved a Million-Dollar Math Problem (New Scientist). An explainer on the Navier–Stokes story and the questions it raised.
AI Is Breaking Math. Will It Solve Physics Next? (PBS Space Time). A wider look at what AI-driven proofs mean for science.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. Which famous problem got a Lean proof developed with Claude in November 2026?
Found a bug? Let us know!
2. What did Claude do with Fermat’s Last Theorem in September 2026?
Found a bug? Let us know!
3. Do the October 2026 ETH claims solve P vs NP?
Found a bug? Let us know!
4. In August 2026, what share of zeta zeros was proved to lie on the critical line (lower bound)?
Found a bug? Let us know!
5. Why does Clay still list the Riemann Hypothesis as unsolved?
Found a bug? Let us know!

Frequently Asked Questions

What were the biggest AI math breakthroughs of 2026?

A Lean proof of the Riemann Hypothesis developed with Claude (November), a Navier–Stokes blow-up result with external force (September), proofs of the cycle double cover conjecture and a Jacobian conjecture counterexample (July), the disproof of Erdős’s unit distance conjecture (May), and a full Lean formalization of Fermat’s Last Theorem (September).

Has AI solved a Millennium Prize Problem?

AI contributed to 2026 results on two of them. OpenAI announced a Navier–Stokes blow-up result in September, which Clay said appears to settle that problem, and a Lean proof of the Riemann Hypothesis developed with Claude was certified by the Lean community in November. No prize had been announced as of November 22, 2026.

Did AI prove the Riemann Hypothesis?

A proof written in Lean and developed with Anthropic’s Claude was released by Levent Alpöge on November 20, 2026, and the Lean/Mathlib community certified the statement and proof on November 21. A human-readable explanation and Clay’s prize process were still pending.

Did AI solve Navier–Stokes?

In September 2026 OpenAI announced an AI-found proof that smooth 3D flows with a smooth external force can blow up in finite time, a case Clay’s problem allows. A November claim for flows with no force was announced, but its public verification package was incomplete.

Did AI solve P vs NP in 2026?

No. In October 2026 OpenAI and Anthropic claimed refutations of the Exponential Time Hypothesis, a stronger assumption than P ≠ NP, and those claims were still under examination. P vs NP remains open.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "AI Math Breakthroughs 2026: What’s Proven So Far - Effortless Math"?

No one replied yet.

Leave a Reply