Claude proves 123 code libraries crash-free

Anthropic's Claude wrote math proofs that 239 entry points in 123 free code libraries cannot crash, Anthropic said on the 22nd of October 2026. An entry point is where outside data enters the code.

The work came from Anthropic's Frontier Red Team with the company Theorem, the research post says. Claude found "hundreds of bugs", fixed them, then proved the fixed code, Anthropic wrote.

123 libraries from Google's OSS-Fuzz, any input up to 64 MB

The libraries come from Google's OSS-Fuzz, which tests free code with random inputs, the post says. Each proof covers every input up to 64 MB, Anthropic wrote.

Some bugs "had eluded detection for years", the post says.

180 million lines of proof, and cost holds it back

The proofs fill about 180 million lines of Rocq, a program that checks proofs, the post says.

Some proofs took hundreds of millions of tokens, the small text pieces AI use is billed by, Anthropic wrote. Lower cost is key to wider use, the company said.

Opus 5.5 used about 40% fewer tokens than Opus 5

On 36 small test tasks, Anthropic's Claude Opus 5.5 proved half with about 40% fewer tokens than Opus 5, the post says. The post does not name the model behind the main run.

Below we break down what a proof means for a normal app user

Normal testing tries many inputs and hopes to hit bad ones. A proof checks all inputs at once, so no input up to 64 MB can crash that code, the post says.

The proofs also cover memory errors, where code reads or writes the wrong spot, Anthropic wrote. For an app user, the proven parts of these libraries should not crash on bad data, within the setup the proofs assume.

What is confirmed and what is not

Anthropic wrote that a formally verified program can still have bugs. The proofs assume one setup: Linux on RISC-V, a kind of chip design, running one task at a time, the post says.

Most people run these libraries differently, Anthropic wrote. The post gives no exact bug count or dollar cost.

Key facts

Anthropic published the research on the 22nd of October 2026.

Anthropic's Claude proved 239 entry points in 123 libraries free of crashes.

It found and fixed hundreds of bugs.

Opus 5.5 used about 40% fewer tokens than Opus 5 on small tasks.