Claude solves 67.2% of the Riemann Hypothesis — AI crosses the assisted-proof threshold

On August 9, 2026, Anthropic published a research note that landed like a silent detonation in the mathematical world. An experimental version of Claude pushed the resolution rate of the Riemann Hypothesis from 41.6% to 67.2% — a 25.6-point leap in a few weeks of computational work. The post immediately hit 122 points on Hacker News. The signal is unequivocal: AI is no longer just assisting mathematicians. It is producing proofs.

This is not another benchmark. The Riemann Hypothesis is one of the seven Millennium Prize Problems of the Clay Mathematics Institute — an unproven conjecture dating back to 1859, whose resolution would trigger a one-million-dollar prize and upend our understanding of the distribution of prime numbers.

What Claude actually achieved

Anthropic’s experiment does not claim a complete proof of the hypothesis. The framework is more nuanced — and more interesting. The researchers decomposed the Riemann Hypothesis into hundreds of sub-problems representing special cases, intermediate lemmas, and equivalent reformulations. Claude was trained to attack each sub-problem autonomously, generating proof proposals that were then formally verified by a proof assistant — most likely Lean 4.

The result: 67.2% of sub-problems solved, up from 41.6% for the previous model version. The gain does not come from a larger architecture or more parameters. Anthropic attributes the improvement to two methodological innovations:

• A tree-search mechanism that lets Claude explore multiple branches of mathematical reasoning before committing to a direction
• An optimized training curriculum for formal proof tasks, where the model learns to distinguish a plausible intuition from a verifiable demonstration

The critical point: Claude’s proofs are mechanically verified. This is not a model that « thinks » it is right — it is a model whose outputs pass a formal checker. The difference is fundamental.

Why the Riemann Hypothesis is a perfect testbed for AI

The Riemann Hypothesis asserts that all non-trivial zeros of the Riemann zeta function have real part equal to 1/2. Since 1859, no one has been able to prove — or disprove — it. The problem has resisted generations of mathematicians, including figures like David Hilbert, G.H. Hardy, and Paul Cohen.

What makes this problem particularly well-suited to mathematical AI:

1. It is decomposable. The hypothesis can be fragmented into range verifications, explicit bounds, and special cases of L-functions.
2. A massive corpus exists. Decades of partial attempts, auxiliary lemmas, and related results are available in the literature.
3. Verification is binary. A proof is either valid in Lean, or it is not. No gray zone.

Anthropic exploited this structure to create a training environment where Claude can iterate rapidly: propose an approach, attempt a formal proof, receive boolean feedback, retry.

The mathematical AI landscape in August 2026

Anthropic is not alone in this field. Competition for AI-assisted proofs is intensifying:

• OpenAI unveiled Astra on August 2, 2026, a model designed for long-horizon reasoning that produced ten breakthroughs in mathematics and theoretical computer science — including the resolution of the Connes rigidity conjecture — at a token cost of $2,000. Its development was partially suspended on August 8 for safety reasons.
• Google DeepMind continues to expand AlphaProof, its automatic proof system combining Gemini with an AlphaZero-inspired tree-search engine.
• Meta is working on hybrid neuro-symbolic approaches for theorem generation.

But Anthropic’s announcement stands out for its methodological transparency. The post details the experimental protocol, the problem decomposition, and the verification metrics — an approach that contrasts with the sector’s more opaque publications.

The 67.2% score means Claude is now closer to full resolution than to the starting line. If the pace of progress holds — 25 points of gain in a few weeks — the 90% bar is conceivable by year-end. Full proof remains an unguaranteed goal, but the horizon is drawing closer.

What this changes for mathematical research

The impact goes beyond the specific case of the Riemann Hypothesis. Three structural implications emerge:

1. Formal verification becomes the bottleneck. Claude can generate thousands of plausible conjectures per hour. The limit is no longer mathematical creativity — it is the capacity to formally verify each proposal. Proof assistants like Lean, Coq, and Isabelle become critical infrastructure.

2. The mathematician’s role shifts. Humans do not disappear — they change position in the chain. Instead of producing proofs, they design problem decompositions, select promising branches, and interpret results. The key skill becomes the ability to ask the right questions, not find the answers.

3. The boundary between conjecture and theorem thins. When a system can formally prove 67% of a problem, the remaining 33% becomes an engineering target — more compute, better decomposition, faster iteration. The question is no longer « is it provable? » but « what compute budget do we allocate to this proof? »

The limitations Anthropic’s post does not hide

Anthropic is honest about current limitations. Claude does not « understand » mathematics in the human sense. The model remains dependent on the quality of the initial decomposition provided by the researchers. If a sub-problem is poorly formulated or rests on a non-formalizable intuition, Claude stalls like any search system.

Moreover, formal verification in Lean imposes a substantial formalization overhead. Translating informal mathematical reasoning into verifiable Lean code remains an art in itself — and Claude is not yet able to cross that bridge autonomously. Anthropic’s researchers had to manually formalize part of the framework before Claude could operate.

Finally, the leap from 41.6% to 67.2% says nothing about the relative difficulty of the remaining 32.8%. It is possible — likely, even, for a problem of this magnitude — that the last sub-problems are exponentially harder than the first ones.

Verdict: what technical teams should take away

Anthropic’s Riemann experiment is not an isolated feat. It fits into a broader trend where AI moves from coding assistant to producer of new knowledge. Here is the conditional verdict:

• If you lead an applied mathematics R&D team: start integrating formal proof assistants into your pipeline. The ability to mechanically verify AI proposals becomes a direct competitive advantage.
• If you invest in scientific AI: the « generative model + formal verifier » pair is the winning architecture of 2026. Generation without verification produces noise; verification without generation hits a ceiling.
• If you follow the frontier model race: mathematical benchmarks — MATH, GSM8K, MiniF2F — are no longer sufficient to measure real capability. The next relevant metric is the resolution rate on non-trivial open problems, not textbook exercises.

Claude did not prove the Riemann Hypothesis. But it demonstrated that a language model can contribute in a verifiable, incremental way to one of the greatest mathematical problems in history. That is a first.

References

• Anthropic — Learning more about Claude’s mathematical capabilities (August 9, 2026)
• Hacker News — Claude moves bound of the Riemann Hypothesis from 41.6% to 67.2% (August 10, 2026)
• Clay Mathematics Institute — Riemann Hypothesis
• Lean 4 — Theorem Prover
• OpenAI — Astra model for long-horizon reasoning (August 2, 2026)