Anthropic Publishes Three Self-Measured Indexes on How Much Claude Leads Its Own R&D

The ruler is Epoch AI‘s. The readings are Anthropic’s own, verified by no external party. That split is the whole question.

What Happened

On 17 September 2026, Anthropic published three measurements of its own AI research and development operations: how much of that R&D is now performed by AI, how well the actions of AI agents are overseen, and how compute is allocated between safety and other work. Every figure in the disclosure is Anthropic measuring itself and publishing the result — no external party has verified any of the ratings.

The automation figures use the Automation Level scale developed by Epoch AI, an independent nonprofit. Anthropic adopted that scale rather than writing its own. Its six points run from AL0 (no AI involvement) through AL3, where AI “collaborates” — completing large chunks of work under close human direction — to AL4, where AI “leads,” completing most of a task end-to-end from a high-level prompt with human supervision, and finally AL5, fully autonomous with no human in the loop. As of August 2026, Anthropic reports that Claude leads 26% of its AI R&D work at AL4 — that is, completing most of a task end-to-end from a high-level prompt with human supervision — and that above 90% of measured tasks now sit at or above AL3. As of February 2026 (an upper bound, not a point estimate), the share Claude led was under 1%.

The same document states, without qualification: “Claude is not operating fully autonomously for any measured subset of AI R&D work.” On oversight, Anthropic reports roughly 30,000 agents running simultaneously as of August 2026, with online monitors covering 100% of agent actions before execution, a blocking rate of 0.002% of decisions (approximately one in 47,000), and offline monitors flagging one to two transcripts per thousand for further review. On compute — measured over the single week of 13 to 20 July 2026 only — 6% of AI R&D compute went to safety work; within AI-driven AI R&D specifically, that share was 12%.

The Structural Read

Four things in this disclosure deserve precise treatment, because each admits a misreading that travels faster than the correction.

First: the number and its boundary. AL4 is not AL5. “Completes most of a task end-to-end from a high-level prompt with human supervision” contains the word “supervision,” and that word does as much structural work in the sentence as the percentage does. A reader who encounters 26% without the AL4 definition has been handed a materially different — and larger — claim than the one Anthropic published. And again: every figure here is Anthropic measuring itself, reported on no external party’s verification.

Second: the zero is the most informative figure. Claude is not operating fully autonomously for any measured subset of AI R&D work. In a week when the industry debated whether frontier development should be paced by agreement, a laboratory published that its own model leads a quarter of its research and that none of that research runs without a person in the loop. Those are two halves of a single disclosure and they belong together wherever the 26% is repeated. It is also necessary to be precise: a share of tasks at a defined automation level is not the same claim as recursive self-improvement occurring or not occurring, and nothing in the document asserts either.

Third: the oversight numbers admit two readings the published data does not separate. A blocking rate of 0.002% — roughly one in 47,000 agent decisions — is consistent with agents rarely attempting anything that warrants blocking. It is equally consistent with a monitor that intervenes rarely by design. The published figures do not distinguish between those two explanations. This piece does not attempt to adjudicate them. The rate is neither high nor low here; no inference is drawn about agent behaviour or monitor quality.

Fourth: the compute split and the third-party commitment. Over one week in July 2026, 6% of AI R&D compute went to safety work; within AI-driven AI R&D specifically, 12% did. The second figure is higher than the first. That is the observation, and it stops there — a single seven-day window is a week, not a trend. The structurally interesting item sits elsewhere: Anthropic states a plan to embed independent third-party evaluators from multiple organisations with access comparable to internal risk assessment teams. That is a stated plan, not an existing fact. It is also the only answer available to the obvious objection about a self-published measurement — because the Epoch AI scale is already external, what an outside evaluator would add is independent application of it. A self-rated measurement invites the question of who checked it, and embedding outside evaluators is the only answer that does not require a regulator.

The Epoch AI Scale Anthropic Adopted

Scale developed by Epoch AI, an independent nonprofit. Adopted by Anthropic. Ratings are Anthropic’s own, unverified by any external party.

Three Implications