Anthropic says AI systems now speed up its own model development
Anthropic reports that AI systems now handle a growing share of its model development, speeding its own progress. The Anthropic Institute analysis, first published in June 2026, draws on public benchmarks and internal data. Its headline figure is that engineers on average ship eight times as much code per quarter as they did from 2021 to 2025. The company adds that more than 80% of code merged into its codebase as of May 2026 was written by Claude, up from low single digits before Claude Code launched in February 2025.
The report traces a path from early chatbots that suggested code snippets, to coding agents that edited whole files, to autonomous agents that take on larger parts of a workflow. Each step reduced the share of engineering that people do by hand. Anthropic says it has not reached fully autonomous recursive self-improvement, the stage where a model designs its successor without human direction, and that this outcome is not inevitable. It also warns that full self-improvement could raise loss-of-control risks.
A revised SEC filing dated November 2 says Anthropic's growing use of AI in research and development could soon meet an internal recursive self-improvement threshold, defined under its Responsible Scaling Policy as AI contributing to a doubling of the usual pace of AI development. The company says no current model has crossed that line and is weighing policy updates to keep humans in control.
Why it matters
For a team building an AI company, the practical issue is timing. If model progress compounds the way Anthropic's internal figures suggest, hiring plans and roadmaps built on steady, linear gains will fall behind. Operators should plan for AI-assisted output to move in large steps, which changes how many engineers a plan needs and which tasks stay with people. Investors should test productivity claims in diligence rather than accept them as settled, since the figures are self-reported.
The loss-of-control link makes monitoring and access controls a near-term engineering task rather than a policy topic. Any team that lets agents write code, launch training jobs or change infrastructure needs logging, scoped permissions and tested rollback paths now. Anthropic's filing itself warns that capabilities could advance faster than the safety, alignment and security measures built to govern them, a risk that applies to any lab or startup shipping agentic tools.
Key points
- 1Anthropic engineers on average ship 8x as much code per quarter as they did from 2021 to 2025.
- 2As of May 2026, more than 80% of code merged into Anthropic's codebase was authored by Claude.
- 3Anthropic itself says the 8x lines-of-code figure is almost certainly an overstatement of productivity gains.
- 4Anthropic says it has not reached fully autonomous recursive self-improvement, where a model designs its successor without human direction.
- 5
