Anthropic measures Claude's role in building AI and outlines what full self-improvement would require

Anthropic discussed how its models already help develop AI and what would count as full self-improvement.

Anthropic Institute published an account of how its engineers and researchers use Claude during AI development. The report, covered on November 2, combines internal observations with evidence about what models can do on longer technical tasks. It asks how far AI-assisted work might progress toward a system that can design and develop its own successor. The institute says that full recursive self-improvement has yet to occur. [Anthropic Institute report](https://www.anthropic.com/institute/recursive-self-improvement)

What the internal numbers capture

The report says that more than 80% of code merged into Anthropic's codebase as of May 2026 was authored by Claude. It also says the typical engineer in the second quarter of 2026 merged about eight times as many lines per day as in 2024. Engineers still direct and review the work. The authors explicitly caution that lines of code capture quantity more readily than value and likely overstate the productivity gain. [Anthropic's internal measurements](https://www.anthropic.com/institute/recursive-self-improvement)

The institute also cites a March poll of 130 people across Anthropic research teams. The median respondent estimated a roughly fourfold increase in output with Mythos Preview compared with doing similar work without AI access. That is a self-reported comparison. It adds information about how researchers perceive the effect, while neither isolating the model's contribution nor measuring an industry-wide average. [Anthropic's survey description](https://www.anthropic.com/institute/recursive-self-improvement)

Which steps still depend on people

Writing code is one step in developing a new model. Someone must also decide which research question matters, run and interpret experiments, obtain the necessary compute, and judge whether a proposed successor performs better under appropriate evaluations. A system that accelerates an engineer's assigned work can change development speed even while humans own those decisions. Crossing from assistance to autonomous successor development would require evidence across the chain.

For readers tracking this question, measures such as the share of experiments proposed, executed and interpreted without intervention would answer a different question from lines of code. Documentation of which access, compute and approvals remained under human control would matter as well. Anthropic's figures describe changes inside the lab. Its account of an autonomous successor remains a scenario about what might follow. [Anthropic Institute report](https://www.anthropic.com/institute/recursive-self-improvement)