180 million lines, a 64-megabyte cap

This Decoder page is dated Oct 23, 2026. It retells work by Anthropic and its partner Theorem. The Frontier Red Team let Claude take over the proofs instead of writing a new piece of software. In a few weeks the model produced about 180 million lines of mathematical proof code, drawn from test programs in 123 open-source libraries in Google's OSS-Fuzz project. The page uses an old comparison to say how expensive this used to be: a team of specialists spent 20 person-years formally verifying the small operating-system kernel seL4. The 20 person-years are an example of the old cost. The 180 million lines are the new scale on this page. The two sentences do not become one accepted engineering total.

An independent program called Rocq checks the proofs automatically. Anthropic says the process needs almost no human review. The researchers call it fully autonomous proof generation. The page adds, immediately, that the tools and the formal foundations were built by people at Theorem. Fully autonomous is their name for it. It is not this story's judgment. Claude checked existing code. It did not, on this page, deliver a separate new product.

Along the way Claude ran into hundreds of bugs and was allowed to fix them. The page gives two examples. While verifying a widely used PDF reader, the team found about as many security bugs in one week as are typically reported for that software in a year. In another application, Claude caught a logical error that had gone undetected for more than a decade. Both examples are their account. The page attaches no bug identifiers. The page also says many of the bugs would have slipped past ordinary testing because they appear only under very specific conditions. Those conditions are their explanation. This story does not write them. Even when a complete proof fails in the end, the page says the attempt still surfaces real bugs. How a bug is triggered is not written here.

The limits are on the same page. Anthropic acknowledges that formally verified does not mean free of errors. The proofs hold only under strict conditions: a specific processor architecture, single-threaded programs, and inputs under 64 MB. In practice these libraries run on other hardware and in other configurations. A single proof can consume hundreds of millions of tokens. No total cost is published. A smaller test also says Opus 5.5 solves half the proofs with about 40 percent fewer tokens than Opus 5. The 40 percent is a comparison for that half. It is not a bill for the whole project.