For years, the intersection of large language models and mathematics has been defined by a frustrating paradox. While AI can solve complex competition-level problems, it remains prone to the subtle, confident hallucinations that are fatal in the world of rigorous proof. The mathematical community has long sought a bridge between the intuitive, probabilistic leaps of an LLM and the absolute, binary certainty of a formal proof assistant. This week, that bridge became a highway as Anthropic demonstrated that its models could not only understand a landmark mathematical proof but translate it into a machine-verifiable format that leaves no room for doubt.

The Scale of Autonomous Formalization

Anthropic recently completed the formalization of Fermat's Last Theorem using an internal general-purpose research model, identified as Claude Fable 5.1. The project was not a manual effort but a largely autonomous operation that spanned 11 days. To achieve this, the system consumed approximately 6 billion output tokens, reflecting the sheer volume of trial, error, and refinement required to satisfy a formal compiler. The final output is a staggering 13 million lines of Lean code, a proof assistant language designed to ensure mathematical correctness. To put this scale into perspective, the resulting codebase is more than five times larger than Mathlib, the standard community-driven mathematical library for Lean.

The architecture relied on a multi-agent harness where dozens of Claude agents collaborated in a tiered pipeline. These agents worked together to prove a total of 30,030 individual theorems, of which 29,500 were ultimately integrated into the final proof. The verification process was rigorous, utilizing the Lean compiler and a specialized comparator tool to ensure that every generated theorem aligned perfectly with the definitions established in Mathlib. The final result was not merely a software output but was personally compiled and verified by Kevin Buzzard, a prominent formal mathematician, confirming that the logic holds under the strictest scrutiny.

The Efficiency Gap and the Prove2Me Catalyst

Despite the logical triumph, the project revealed a stark contrast between machine-generated correctness and human-centric elegance. The 13 million lines of code are characterized by extreme redundancy and inefficiency. Unlike Mathlib, which is meticulously curated by humans to be concise and maintainable, Claude's output is bloated. This lack of optimization manifests in severe performance bottlenecks: even on a high-end machine with 96 cores, the compile time for this proof is approximately 20 times slower than that of Mathlib. The memory footprint is equally oppressive, making repository navigation cumbersome even on systems equipped with 500GB of RAM. This creates a critical insight for the AI community: achieving logical accuracy is a separate challenge from achieving software quality.

The breakthrough that enabled this scale was the integration of Prove2Me, a platform developed by researchers including Tianyi Peng at Columbia University. In early attempts, the Claude agents failed to collaborate because they lacked a shared understanding of the project's global state. This coordination failure resulted in a significant amount of wasted effort, with failed code accounting for roughly 7% of the total output. Prove2Me solved this by maintaining a dependency graph of theorems. By separating theorem statements from their proofs into distinct files, the platform increased Lean's compilation speed and reduced resource consumption. Each theorem was tagged with a natural language description, allowing agents to search for existing results or identify simpler paths to a proof. Following the 1995 proof established by Andrew Wiles and Richard Taylor, Claude navigated the complex logic of the Langlands-Tunnell theorem and Ribet's level lowering theorem to reach the final conclusion.

This achievement is less about discovering new mathematics and more about the industrialization of formalization. The ability to convert thousands of pages of existing mathematical literature into machine-verifiable code in a matter of days suggests a future where the bottleneck of human verification is removed. This potential for accessibility was further proven in a smaller-scale experiment where the Vinogradov three-prime theorem was formalized in just three days using only three personal Claude Max accounts and the Prove2Me platform.

The integration of formal verification tools like Lean into the AI research pipeline provides a mechanism to instantly validate hypotheses and intermediate results. As AI continues to generate vast quantities of mathematical data that exceed the capacity of human peer reviewers, the emergence of a hybrid standard—where a human-readable paper is accompanied by a computer-verified formal proof—will likely become the new benchmark for mathematical truth.