Anthropic published a first-of-its-kind look inside its own development pipeline this week, disclosing that its Claude model now "leads" 26% of the AI research and development work at the company — up from under 1% in February 2026. The figures appeared in a company blog post titled "Measurements for understanding the pace of AI development inside frontier labs," published by Anthropic on September 17 and reported by Reuters, The Washington Post, and other outlets.
The post is partly a transparency exercise and partly an argument: Anthropic wants other frontier labs to publish the same kinds of numbers, so that the public and governments can track how quickly AI is being handed the keys to building AI. For more context on this story, see our ongoing latest AI developments.
Three Numbers to Track AI Development
The company proposed three measurements that it says illuminate the pace of development inside frontier labs: how much of AI R&D is performed by AI itself, how well the actions of AI agents are overseen, and how compute is allocated between capability work and safety work.
The headline number comes from what Anthropic calls its R&D Automation Index. The company catalogued every kind of AI R&D work done at the company, rated how automated each task currently is, and aggregated the results using a scale developed by the research group Epoch AI. The scale runs from AL0 (no AI involvement) to AL5 (AI operates fully autonomously, with no human in the loop), with AL3 meaning AI "collaborates" on large chunks of work under close human direction and AL4 meaning AI "leads," completing most of a task end-to-end from a high-level prompt while a human supervises.
As of August 2026, Anthropic reported, Claude leads 26% of the company's AI R&D work. The share of work at or above the "AI collaborates" level is above 90%. Claude is not operating fully autonomously for any measured subset of AI R&D work.
30,000 Agents, One Monitor
The second measurement concerns oversight. Anthropic disclosed that as of August 2026 there were approximately 30,000 agents doing research and engineering work at the company at any one time on its most-used internal platform.
Every action those agents take passes through an online monitor before it is executed, usually within seconds, according to the post. Anthropic says it analyzed over a billion decisions from its research and engineering agents over August and found that the share of activities blocked or escalated by the monitors was 0.002%.
On compute allocation, the third measurement, the company said that over the examined week roughly 6% of the compute that went to AI R&D was allocated toward safety.
Building Its Own Successor — With Humans Supervising
The framing matters as much as the figures. Anthropic's models are increasingly participating in building the next generation of those same models — what researchers call recursive self-improvement, and what The Washington Post's headline described more bluntly as a chatbot "taking over the work of building its own successor."
Anthropic is careful to draw a line between "AI leads" and "AI replaces." At AL4, a human still supervises and can intervene; the company explicitly reported that no measured subset of its R&D work runs at AL5. But the trajectory is steep: going from under 1% to 26% in roughly seven months suggests the share of AI-led work could pass half well before the end of the decade if current trends hold.
The methodology has real limits, and Anthropic says so. Its automation ratings were produced by sampling the company's own work records — including Slack messages and internal documentation — and having both humans and a judge model rate how automated each task is. In a blind test, staff rated tasks without knowing what evidence the models had gathered. The company reported that its model judge agreed with humans about as often as humans agreed with each other: exact model-to-human agreement was 59%, while human-to-human agreement was just 35%, and model and human ratings landed within one level of each other 97% of the time.
The company also acknowledged the self-referential risk in its own methodology: Anthropic used its own models to help evaluate its systems, which could mean the judge model makes the same kinds of errors as the model it is checking. Third-party verification, it argues, is the real answer.
Outside Eyes Coming
To that end, Anthropic says it plans to embed independent third-party evaluators from multiple organizations inside the company, giving them access to internal processes, systems, and data comparable to what internal risk assessment teams have. The external AI research nonprofit METR has previously red-teamed Anthropic's offline monitoring platform.
The disclosure arrives amid an escalating debate about pacing. Anthropic CEO Dario Amodei has called for coordination on slowing the frontier, and the company notes that its own numbers would be expected to shift if such coordination existed. This week, Anthropic and OpenAI were also the subject of a public letter from safety experts arguing that the labs need truly independent safety evaluators.
Whether rivals follow Anthropic's lead is the open question. The company argues any frontier developer could publish these measures regularly using a public methodology. Until they do, the only public numbers on how fast AI is building AI come from the labs themselves.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
