AGENTTechnology and AI
← Latest editionClaude Is Now Doing a Quarter of Anthropic’s AI Research
Anthropic says Claude leads 26% of its model-building work, even as the company calls for a slower frontier race.
- Claude “leads” 26% of Anthropic’s AI research and development work, according to the company’s own measurement for August 2026. 1
- Anthropic defines “leads” as completing most of a task from a high-level prompt while a human supervises. Claude is not yet fully autonomous on any measured work. 1
- More than 90% of the company’s AI R&D is now at least conducted in collaboration with Claude, up from less than 1% at the “leads” level earlier this year. 1
- The figure comes from Anthropic’s internal records, a sample of staff activity and an independent Claude model judging the evidence — not from an outside audit. 1
- Anthropic says about 30,000 agents were carrying out research and engineering work on its main internal platform in August, with roughly one in 47,000 monitored decisions blocked. 1
The important change is not that Claude has built a successor to itself. It has not. Human researchers still choose the problems, set the goals and approve the results. But a growing share of the work between those decisions — writing code, running experiments and investigating failures — is being handed to systems made by the company that benefits from moving faster. 1
That makes Anthropic’s call for a slowdown politically awkward. The company says the public needs better information about how quickly frontier labs are automating AI development, and argues that advanced development may need to pause or slow. At the same time, its own data describes a race in which the work of building better models is becoming easier to automate. 2
The measurement is useful evidence, but it is not an independent verdict. Anthropic built the task catalogue from internal Slack and documentation, used Claude to classify much of the work, and acknowledges that there is room for disagreement about where “collaborates” ends and “leads” begins. It is asking other labs and outside evaluators to adopt comparable methods — because, for now, the companies building these systems remain the people measuring how close they are to losing control of the process. 1