An Anthropic formal-math repository claims that an algorithm discovered by an internal research Claude model refutes the Exponential Time Hypothesis, or ETH. Its README says the deterministic algorithm solves 3-SAT on n variables in time 2^o(n). It also claims a bound of 2^O(n log log n / √log n) for counting satisfying assignments, a problem the README describes as harder.
The same README puts a firm limit on the practical takeaway: “The result is purely asymptotic and says nothing about solving formulas in practice.” The directory lists a paper credited to Claude (Anthropic) and links to Lean 4 files. Its Oct. 27, 2026, commit is labeled as adding the paper and formalization.
What the proof files are meant to check
The repository's explanation of its Lean proof says MainResults.lean states both that ETH is false and the paper's main counting-algorithm theorem. It provides build instructions and describes a fuller check using a second proof-checking system. But the README also draws a distinction: Lean checks proofs, while “no machine checks that this file says what the paper says.” That correspondence remains a question for a reader to check, not something established by the listed commands alone.
A separate Oct. 27 post by @alpoge quotes a statement saying the core algorithm was found Oct. 16 using 14 million output tokens and no human input. The post says Anthropic researchers then worked with Claude on the paper's presentation. The Oct. 16 date, token details and account of human involvement do not appear in the repository's README, and the linked GitHub page for the PDF displays a file wrapper rather than the paper's text. Reacting to the post, @AndrewCurran_ called the claim “extremely confusing.”
Comments