Anthropic reports Claude writes over 90% of its code and calls for a verifiable global AI development pause

Anthropic published internal data showing engineers in Q2 2026 ship an average of eight times as much code per day as in 2024, with more than 80 percent coming from Claude after its Code feature launched in February 2025. Leadership estimates the total share—including scripts and experimental code—at more than 90 percent.

Claude's ability to handle longer autonomous tasks is accelerating: task duration has doubled roughly every four months (down from seven), with Claude Opus 4.6 now reliably tackling 12-hour tasks. METR found Claude Mythos Preview could work for 'at least' 16 hours. In research optimization, Claude Mythos Preview reached roughly 52x speedup in May 2026, versus about 3x for Claude Opus 4 a year earlier.

Anthropic acknowledges the eight-fold code increase 'is almost certainly an overstatement of the true productivity gain,' citing a March 2026 internal survey of 130 employees that pegged the median estimate at 4x. Claude-written code was 'somewhat worse than human-written code' in late 2025, 'is roughly at parity today,' and is expected to be 'strictly better within the year.'

Anthropic states it would slow or pause its own frontier AI development if other developers at or near the frontier did the same in a verifiable way, and the Anthropic Institute plans to research systems to make such a pause credible. The company notes that training runs are 'far easier to hide than missile silos' and warns 'whoever continues while others pause could inherit the lead.'