Claude Now 'Leads' 26% of Anthropic's Own AI Research — Inside the R&D Automation Index

On September 17, 2026, Anthropic published something frontier AI labs have never released before: hard, methodology-backed numbers on how much of the work of building AI is now done by AI itself. The centerpiece statistic is startling. As of August 2026, Claude “leads” 26% of Anthropic’s own AI research and development — up from under 1% in February — and the share of R&D work at or above the “AI collaborates” level exceeds 90%. The model is no longer just a coding assistant at the lab that built it; it is performing a quarter of the research function.

The numbers arrive in a post titled “Measurements for understanding the pace of AI development inside frontier labs,” published under the Anthropic Institute banner with co-authors Marina Favaro and Phillie Wright, research direction from co-founder Jack Clark, and technical proofs of concept from a team including Jun Shern Chan, Fabien Roger, and Holden Karnofsky. The explicit framing is policy-facing: as governments and the public debate slowing the pace of frontier AI development, Anthropic argues, the public needs visibility into what is actually happening inside the labs. “We should do everything possible to minimize the gap between what frontier labs know and what the public knows,” the company writes.

Three measurements, one snapshot

The post lays out three metrics intended to become a recurring, verifiable reporting standard.

The R&D Automation Index. Anthropic catalogued every kind of AI R&D work performed at the company, rated how automated each task currently is, and aggregated the ratings into a single index. The rating scale is not homegrown — it adopts the “Automation Level” (AL) ladder proposed by Epoch AI, running from AL0 (no AI involvement) to AL5 (fully autonomous, no human in the loop). In between sit the two levels that matter most: AL3, where AI “collaborates” — doing large chunks of work under close human direction — and AL4, where AI “leads” — completing most of a task end-to-end from a high-level prompt while a human supervises.

The August 2026 reading: zero measured work at full autonomy, 26% at “leads” or above, and over 90% at “collaborates” or above. In other words, Claude is nowhere near building its own successor unsupervised, but the trajectory from roughly zero in February to a quarter of R&D by late summer is the steepest part of the story. Anthropic notes it would expect these numbers to shift if there were coordinated pacing of the frontier, as CEO Dario Amodei has called for.

Oversight of agents. Roughly 30,000 AI agents were doing research and engineering work at Anthropic at any one time, on the company’s most-used internal platform alone. Every one of their actions passes through an online monitor in real time before execution — usually within seconds — designed to stop fast, irreversible harms like an agent copying model weights out of the system. Offline monitors then read agent activity after the fact, hunting for slowly developing problems such as signs of misalignment. The company analyzed over a billion decisions from these agents during August 2026 and found that 0.002% — about 1 in 47,000 — were blocked. The monitor flags roughly 100,000 transcripts per week for classifier review, with the highest-priority ~50 per week escalated to humans. METR has independently red-teamed the offline monitoring platform in the past, and Anthropic says it is now standing up embedded third-party evaluators with access comparable to internal risk teams.

Compute allocation. During a snapshot week (July 13–20, 2026), about 6% of the compute going to AI R&D was allocated to safety work, and about 12% of the compute going to AI-driven AI R&D was safety-focused. Anthropic stresses these are deliberately conservative: work that advanced capabilities as much as safety was excluded, and safeguard classifiers’ compute is counted separately.

How you measure “AI building AI” without fooling yourself

The appendix is the most interesting part of the document, because it confronts the obvious objection head-on: Anthropic used Claude to measure Claude. The methodology is a piece of measurement engineering in itself.

No single person can enumerate every R&D task at a frontier lab, so Anthropic built the task list bottom-up. For each week of July 2026 it randomly sampled 20% of staff in every department in the model R&D loop, then had a Claude research agent read each sampled person’s week — Slack messages and internal documentation — and list the tasks they worked on. That produced roughly 15,000 granular tasks, which Claude then organized into a hierarchical tree of 542 nodes, 378 of them leaves with names like “eval platform defect diagnosis and fixes” and “serving incident postmortems.” The tree is frozen, so every future measurement is scored against the same basket of work. Each node’s weight is person-time: every sampled person contributes one unit per week, split evenly across their tasks.

To check whether the judge model could be trusted, Anthropic asked staff who own each work area to rate their areas’ automation blind — without seeing what evidence the models had gathered. The results are remarkably candid: model-versus-human exact agreement was 59%, while human-versus-human agreement was only 35%, and model and human ratings were within one automation level of each other 97% of the time. The company also verified that the structure of R&D work isn’t simply shifting under the frozen basket: comparing monthly task arrivals from February to July 2026 against a January baseline found no rise in “novel” tasks.

Why it matters

The significance is less the specific numbers than the precedent. Recursive self-improvement — a model autonomously building its successor — has been a theoretical threshold discussed in safety literature for years. The Automation Index turns it into a tracked, auditable quantity with a public methodology that, in Anthropic’s words, “any frontier developer could publish.” The company explicitly invites cross-lab comparison while acknowledging two obstacles: no shared methodology exists yet, and self-evaluation risks the judge model sharing the blind spots of the model it is judging. Its proposed remedies are third-party verification, cross-model judging with guardrails for competitively sensitive data, and potentially using such metrics as triggers for stronger requirements — like a fixed testing window before a new model is used for further AI R&D.

The context makes the move strategic as well as scientific. Anthropic ties the measurements directly to its proposed Advanced AI Framework (AAIF), which would impose transparency obligations on any lab releasing frontier models, and to Amodei’s calls for coordinated pacing. Publishing unflattering-adjacent numbers — a 6% safety compute share will be quoted back at the company — is a bid to set the reporting standard before a regulator sets it for them. Reuters and Bloomberg both led with the 26% figure; expect “AL4 share” to become a standard question in earnings calls and policy hearings about every frontier lab.

The honest caveats are in the document itself: one week of compute data proves the measurement can be made, not a trend; the frozen basket tracks automation of July 2026 work, not new work humans may have migrated to; and compute share measures spending, not effort — a more efficient safety classifier mathematically shrinks the safety percentage without any safety work being cut. But as a first public instrument panel for recursive self-improvement, it is a genuine addition to how the outside world can watch the most consequential industrial process on the planet.