As of August 2026, Anthropic estimates that Claude was leading 26% of the AI research and development work covered by its internal measurement system. In March, the figure was roughly 1%.

That does not mean Claude is independently training its successor. Anthropic explicitly says none of the measured work is fully autonomous. Humans still set objectives, supervise systems and review results.

What has changed is the level at which Claude can operate. Instead of simply producing a code snippet after receiving a tightly specified instruction, an agent can increasingly receive a broader objective, devise an approach and carry out a substantial portion of the work before asking a human to intervene.

Claude is involved in almost everything

Anthropic says more than 90% of its measured AI R&D work in August included at least some human-AI collaboration. That makes Claude less of an occasional productivity tool and more of a permanent participant in the development process.

The company uses an automation scale based on work from Epoch AI to distinguish between assistance, collaboration and AI-led work. Anthropic stresses that its systems have not reached the final level of full autonomous R&D.

The coding numbers already show how quickly the workflow has shifted. Anthropic says more than 80% of the code merged into its codebase was authored by Claude as of May 2026, with engineers increasingly directing, reviewing and integrating the work rather than writing every line themselves.

Thirty thousand agents running at once

The internal deployment is large. Anthropic reports that approximately 30,000 AI agents were performing research and engineering tasks at any given time on its most-used internal platform in August.

Their actions are screened before execution by an online monitoring system. Across more than one billion agent decisions examined during the month, Anthropic says roughly one in 47,000 was blocked.

Offline monitoring then examines activity after the fact and escalates selected cases for human review. The point is not that the failure rate has disappeared. It is that Anthropic is trying to build an oversight layer capable of scaling alongside a rapidly growing population of autonomous workers.

The feedback loop is the important part

If AI makes AI research faster, the next generation of models can arrive sooner. If those models are better at research, they can accelerate the following generation again. Anthropic describes the endpoint of that hypothetical cycle as recursive self-improvement.

The company says it has not reached that point, and it does not claim that fully autonomous AI research is inevitable. The current data does show the first portion of the loop becoming measurable.

Anthropic engineers are also shipping far more code than they were before Claude became deeply integrated into development. According to the company's research, output per engineer has risen sharply since models began executing longer tasks rather than merely suggesting code.

Measuring acceleration before it becomes opaque

Anthropic wants other frontier labs to publish similar metrics. Its proposal includes tracking the share of R&D led by AI, the number of active agents, how comprehensively those agents are monitored and how much compute is devoted to safety.

In one sampled week in July, Anthropic says about 6% of compute devoted to AI R&D went toward safety work. Within AI-driven research specifically, the figure was about 12%.

There are methodological caveats. Drawing a clean line between an AI system assisting, collaborating and leading a task is inherently fuzzy, and Anthropic's measurements include automated classification. The 26% figure is therefore more useful as a trend indicator than as an exact boundary.

The trend itself is clearer. Claude is no longer merely a product made by Anthropic's researchers. It is becoming one of the tools doing the research that produces the next Claude.