Published: 2026-09-18 | By: Carlos Montiel | Reading time: ~5 min
Anthropic revealed that Claude now leads 26% of its own AI research and development work, a huge jump from under 1% in February 2026. It's a striking headline ("AI building AI"), but Anthropic's fine print is more nuanced than it sounds, and it's worth reading before drawing conclusions.
What that 26% actually measures
The figure covers AI R&D work across the whole company and does NOT mean Claude is independently building its next model. More precisely, that 26% corresponds to the stage where "humans set the goals and supervise, and Claude executes most of the work" — meaning human direction with execution delegated to Claude, not full autonomy.
The rest of the work isn't 100% human either
Claude's involvement levels in Anthropic's R&D:
- Over 90% of measured work reached at least "AL3" level
(active Claude collaboration, not just minor assistance)
- 26% reached the level where Claude executes and humans supervise
- 0% of measured categories operate with full autonomy
- ~30,000 AI agents working in parallel on Anthropic's most-used
internal platform (as of August 2026)
The jump from under 1% to 26% in just seven months is, in itself, the most revealing part of the announcement: it isn't that Claude suddenly acquired a new capability, it's that Anthropic deliberately accelerated how much of its own research workflow it delegates to instances of its own model working in parallel.
Why Anthropic keeps stressing the word "supervision"
Anthropic explicitly stating that "Claude does not operate with full autonomy in any measured category" isn't a throwaway detail — it's the line the company needs to keep clear for its own safety narrative. A company whose central pitch is "we build AI responsibly" can't, at the same time, announce without caveats that its model already builds models with no human oversight. The 26% is real and significant, but it's deliberately framed not to cross that line.
What this means for companies evaluating agentic AI
Beyond the headline, the useful operational detail is this: Anthropic is running ~30,000 agents in parallel, with humans setting goals and Claude executing the heavy lifting, and explicitly classifying the autonomy level of each task (the "AL" framework they mention). That pattern — goals defined by humans, execution delegated to agents, and an autonomy level that's measured and classified, not assumed — is exactly the kind of governance any company should demand of its own AI agent deployments in internal processes, not just of a frontier lab.