Unconfirmed. New Runtime has not found an official statement. This note reports what 4 independent accounts say; treat it as a rumour until an official post, documentation or pricing page confirms it.
According to posts on X, Anthropic worked with a company called Theorem on using Claude to formally verify 123 open-source libraries at the assembly-code level. Several accounts describe the work as producing about 180 million lines of proof script over a period of weeks. The accounts characterize formal verification as a way to prove that code meets specified properties, rather than relying only on tests that try many inputs.
Posts say the effort uncovered bugs that had not been found despite extensive testing, with one account citing trillions of test runs. Other posts caution that Claude was able to modify the code it was proving, and raise questions about how to ensure those changes do not simply remove behavior or failure modes. One post says a Rocq kernel checks each proof and mentions a reward penalty and audit as safeguards; these details are claims made by posters, not independently verified here.
The posts frame the work as a possible expansion of formal verification beyond small codebases or projects requiring years of expert effort. New Runtime has not found an official statement from Anthropic confirming the claims. An official post, research documentation, or other company-published materials describing the project would help establish what was done, what properties were proved, and what safeguards were used.
If it is true: If accurate, the reported work could make rigorous software verification practical for more large codebases; the caveat is that posts also raise unresolved questions about code changes and proof safeguards.