Anthropic: 80% of Merged Code Now Claude-Authored
Anthropic has published internal benchmark data showing more than 80% of all code merged into its codebase is now written by Claude, with engineers shipping 8× more code than in 2024. On open-ended engineering tasks, Claude's success rate jumped from 26% to 76% in 6 months. A training optimization benchmark hit 52× speedup versus a human researcher baseline of 3–4×; Claude proposed a better next research step than the human 64% of the time at mid-session checkpoints.
Why It Matters
Anthropic explicitly states recursive self-improvement is not yet achieved, but the internal compounding rate — now backed by hard numbers — is outpacing public benchmark expectations and is entangled with the company's concurrent IPO filing.