Anthropic says 80% of its new production code is now authored by Claude

Anthropic disclosed that 80% of new production code merged in May was written by Claude, a milestone the company ties to an 8x jump in code shipped per engineer per quarter compared with the 2021-2025 baseline. The claim is notable both as a productivity data point and as a strategic narrative: Anthropic explicitly invokes 'recursive self-improvement,' the idea that AI systems increasingly build the next generation of AI systems.

The mechanics rest on agentic endurance. Anthropic cites long-duration evaluations in which Claude Opus 4.6 sustains coherent work across 12-hour tasks, while its restricted Claude Mythos Preview extends past 16 hours of continuous problem-solving on complex, open-ended engineering problems that lack clear specifications. That endurance — not just raw benchmark scores — is what lets a model own large swaths of a real codebase rather than just autocomplete snippets.

Competitively, the disclosure lands amid an escalating coding-model war. OpenAI shipped GPT-5.3-Codex the same week and is embedding Codex across the ChatGPT app, while xAI says its 1.5T-parameter Grok V9-Medium was trained on Cursor workflows specifically to answer Claude's coding lead. Anthropic's framing — 'how your enterprise can keep up' — is also a sales pitch aimed at the enterprises now weighing Claude against GPT-5.5 and Gemini for their own engineering org.

Skeptics note the figure is self-reported and that 'authored by Claude' bundles boilerplate, tests and refactors that humans still review and merge. The bigger question is whether productivity gains compound or plateau as review and integration become the bottleneck — and what it means for human engineers if Anthropic's own claim that Claude will outperform its staff by 2029 holds.