Claude Formalizes Fermat's Last Theorem

This week's Tuesday Special analyzes a case that bridges two typically distant worlds: pure mathematics and AI verification engineering. The report published by Anthropic on September 4, 2026 documents the first end-to-end, computer-verified proof of Fermat's Last Theorem, produced largely autonomously by Claude over the course of 11 days.

The case of the week: methodology and test origin

The project was led by Tianyi Peng, an Anthropic researcher whose group at Columbia University builds tools for AI-assisted formalization. The initial goal was to verify whether a language model could contribute to coding in Lean an existing proof, the one signed by Andrew Wiles in 1995, spanning 129 pages and manually verified for months by the mathematical community.

The theorem, annotated by Pierre de Fermat around 1637 in the margin of a copy of Diophantus's Arithmetica, states that the equation a^n + b^n = c^n has no solutions among positive integers for any value of n greater than two.

The central finding: exact and verifiable numbers

The evidence reported is quantitative and traceable: in 11 days of work conducted largely autonomously, Claude produced 13 million lines of Lean code and proved 29,500 intermediate theorems, according to documentation by Anthropic^[1].

The resulting code constitutes the first end-to-end proof of the theorem entirely checked by a formal verification assistant.

Kevin Buzzard, lead of the community formalization project launched in 2024 at Imperial College London, called the result an extraordinary milestone in self-formalization, emphasizing that the proof rests solely on the axioms of mathematics.

Why verification matters more than discovery

The distinction between discovery and verification is methodologically relevant for those evaluating investments in AI applied to research. Unlike recent work on the Riemann hypothesis, which produced novel mathematics, the contribution documented here concerns the ability to verify an existing proof by composing it into steps individually verifiable by the Lean system.

Each intermediate theorem represents a link in a logical chain: should a single step prove fallacious, the entire subsequent structure would lose validity.

This mechanism of automatic verification is what makes the result reproducible, unlike a benchmark that measures performance under standardized conditions and lacks direct connection to reliability in production. The gap between a benchmark score and a line-by-line verified proof is where the entire question of trust in AI systems is decided.

Error compounding along long chains of reasoning

The 29,500 intermediate proofs constitute a natural sample for observing what happens when errors propagate along long chains of composite reasoning. Literature on multi-step agent reliability documents a recurring structural pattern: errors rarely sum linearly, they tend instead to multiply.

An agent with 90% accuracy over five consecutive steps achieves 59% composite accuracy on the overall task, a figure that directly concerns systems that chain thousands of steps like the one described here.

In the case of Fermat formalization, the Lean system functions as a control mechanism that intercepts error before it propagates, a verification infrastructure that is absent in the majority of current enterprise deployments.

What it means for Chief Revenue Officers, Chief Data Officers, and Investment Committees

For those sitting on an investment committee, the relevant data concerns where to direct R&D budget allocated to AI applied to complex, long-chain tasks.

The Fermat case shows that prolonged autonomy, 11 consecutive days according to the Anthropic report, becomes manageable when there exists a level of formal verification capable of checking every single intermediate step, rather than merely evaluating the final output.

Capital allocation to multi-step agents lacking an equivalent verification layer remains a methodologically weak hypothesis: the documented result applies to a domain, formal mathematics, equipped with a native verification language like Lean. Extending the conclusion to other domains lacking such infrastructure exceeds what the authors themselves claim in the report.

Data infrastructure for the Chief Analytics Officer

The Chief Analytics Officer reads this case as an infrastructure indicator, even before reading it as a product indicator.

The public repository of the proof, available on GitHub^[2], makes every intermediate step inspectable, a requirement that rarely accompanies AI results communicated to the public.

Building equivalent capabilities in a business context requires a system of granular tracking of intermediate steps, not solely of final output: this is the difference between measuring a pilot's results in a controlled environment and measuring the entire chain that produces it under real-world load.

Which technology thesis holds for the Board

For the board, the Fermat case supports a precise technology thesis: self-formalization can reduce the verification time of complex results, a process that traditionally requires years of collective work by the scientific community.

Kevin Buzzard observes that artifacts produced by self-formalization prove robust enough to be built upon, a condition that makes plausible an acceleration of scientific verification at scale.

The opposing thesis still stands for review, the one treating benchmarks as sufficient proxy for production readiness: the case described here demonstrates the exact opposite, namely that formal verification, and not performance on a standardized test, constitutes the relevant metric for capital allocation to long-chain AI systems.