Claude Now Leads 26% of Anthropic's Own AI Research

Anthropic's new transparency report says Claude led 26% of its model R&D in August, up from under 1% in February, with 30,000 agents at work.

Anthropic published a report on September 17 stating that Claude "leads" 26% of the company's measured AI research and development work as of August 2026. On the automation scale developed by Epoch AI, "leads" means AL4: the model can complete most of a task end-to-end from a high-level prompt while a human supervises, though no measured area is fully autonomous at AL5. Anthropic says more than 90% of its R&D sits at or above AL3, "AI collaborates." The company argues any frontier lab could publish the same three measures — AI-led research share, agent oversight, and safety compute — using a public methodology.

The trajectory is the surprising part. MIXED Reality News reports Anthropic's own chart shows the "leads" share rising from under 1% in February 2026 to 26% in August, while Tekedia describes the climb as from 1% in March to 26% — a sharp jump either way. Hoodline adds a striking projection, attributed via AlphaSignal: if the current pace holds, Claude could be leading up to 80% of measured internal R&D at AL4 by the end of the year. Anthropic frames the disclosure as an attempt to narrow the gap between what frontier labs know about development pace and what the public can see, especially as its own CEO Dario Amodei calls for slowing AI development over safety concerns.

The methodology has a self-referential wrinkle Anthropic openly concedes. According to MIXED Reality News, the company sampled 20% of staff in each relevant department weekly in July, had a Claude agent list their tasks from Slack and internal documents, then had a separate Claude "judge" rate how automated each kind of work is. Anthropic admits the judge "could make the same kinds of errors as the model it is checking," though it says the judge agreed exactly with staff ratings 59% of the time — while staff agreed with each other only 35% of the time. The 26% figure is also self-reported and has not been independently verified, which is why Anthropic plans to embed third-party evaluators with access to its internal systems.

Anthropic says roughly 30,000 AI agents were doing research and engineering work at any one time in August on its most-used internal platform, and the figures cover that platform only. Every agent action is screened before execution: across more than a billion decisions in August, 0.002% — about one in 47,000 — were blocked, with humans reviewing blocked actions within a week. A second, after-the-fact monitor flags roughly 100,000 transcripts a week, of which about 50 reach a human reviewer. Anthropic's own point is that individual agents rarely misbehave, but at millions or billions of agents "even rare events can happen regularly."

The third measure covers compute. In a single sample week, July 13 to July 20, Anthropic says about 6% of the compute used for AI research went to safety work, and about 12% of the compute used for AI-driven research. The company calls these deliberately conservative numbers: work that helped capabilities as much as safety counted as general research, and its safeguard classifiers, which it says use a separate and comparable amount of compute, are excluded entirely. MIXED Reality News notes one week shows the measurement can be made, not a trend, and Anthropic itself cautions that compute allocation is not a perfect proxy for safety effort since safety research can be human-heavy and compute-light.

Anthropic says it plans to publish these measures regularly and to bring in external evaluators from multiple organizations with access to internal processes comparable to its own risk teams. The disclosure landed the same week OpenAI introduced its own reporting framework for model misalignment and published six incident reports, suggesting the industry is moving toward standardized self-measurement. The timing also coincides with Amodei and other tech leaders calling for a slowdown in development, which makes credible, third-party-verified automation metrics more consequential.