Anthropic puts numbers on itself: Claude now "leads" 26% of the R&D that creates the next Claude

By Zendoric · September 22, 2026.

Anthropic has released the first quantified snapshot of how much of the research and development (R&D) on its own artificial intelligence models is already being done by AI itself. According to the company's data, in August 2026 Claude "leads" 26% of AI R&D work at Anthropic: it completes most of a task from start to finish from a high-level instruction, with a human merely supervising. More than 90% of the work now sits at least at the "collaborates" level, in which the AI executes large parts of a task under close human direction. No part of the measured work reaches full autonomy, the company itself specifies.

The figure comes from an "Automation Level" (AL) scale developed by the research firm Epoch AI, running from AL0 (no AI involvement) to AL5 (the AI operates fully autonomously, with no human in the loop). Anthropic uses it to build what it calls its "R&D Automation Index": a catalogue of every task carried out at the company to build models, scored according to how much of that task AI already does.

The release is no accident. Anthropic explicitly frames it within the debate over whether the world should coordinate a slowdown in frontier AI development —what the company itself calls "pacing"—, an idea its chief executive, Dario Amodei, has publicly defended. For such an agreement to be verifiable, something is needed that does not exist today: shared, auditable metrics on how fast each lab is advancing and how much of that progress no longer depends on human hands. Anthropic says it plans to bring in independent external evaluators with access to its internal processes, systems and data comparable to that of its own risk evaluation teams.

The methodology, detailed by the company, is ambitious: each week of July 2026, 20% of the staff in the departments involved in the model R&D cycle was sampled, a Claude agent reviewed their activity on Slack and internal documentation, and the result —some 15,000 granular tasks— was organized into a hierarchical tree of 542 nodes and 378 leaves, from "diagnosing failures in the evaluation platform" to "post-mortems of service incidents". Another model, acting as judge, rated the automation level of each branch.

A caveat is in order here, one Anthropic itself acknowledges: the judge is an in-house model evaluating in-house work. To check for bias, the company compared the model's ratings with those of employees familiar with those areas, without telling them what evidence the model had used. Exact agreement between model and human was 59%, similar to the 35% exact agreement among humans themselves; and model and human agreed within one level of difference in 97% of cases. It is a reasonable validation, but it is still a company certifying its own pace, not a third party.

The second metric measures something different: how much human oversight remains when work is delegated to agents that operate semi-autonomously over extended periods and coordinate with one another. Anthropic puts at around 30,000 the number of research and engineering agents operating simultaneously in August 2026 on its most-used internal platform. 100% of their actions pass through an automatic monitor before being executed —designed to head off irreversible harm, such as the unauthorized copying of a model's weights— and 100% are also logged for later review, aimed at detecting misalignment behaviours that develop more slowly.

Across more than a billion decisions analysed during August, the real-time monitor blocked 0.002% —roughly 1 in every 47,000—. The monitor that reviews after the fact flags one to two transcripts out of every thousand for further review, some 100,000 a week, of which only around 50 reach a human. These are low figures, and the company itself notes that individual agents rarely misbehave; but it also warns of the arithmetic that matters: if millions or billions of agents operate across the economy, even rare events become frequent. The external evaluation firm METR has already carried out independent red-teaming of this monitoring platform in the past, and Anthropic says it is now setting up that kind of external evaluation with third parties.

The third leg is the allocation of compute, the resource —chips, energy, processing time— that underpins the training and operation of the models. In a sample week (13 to 20 July 2026), Anthropic says 6% of the compute earmarked for AI R&D went to safety work, and 12% of the compute earmarked for AI-driven R&D also went to safety. The company itself calls these figures "deliberately conservative": a token that serves both capabilities and safety is not counted, and the compute for safeguard classifiers, of comparable volume, falls outside these metrics.

Our reading is that this exercise is worth more for the infrastructure it creates than for the figures it yields today. 26% of R&D already "led" by Claude, with zero full autonomy, marks a concrete point on a trajectory that until now was only discussed in vague terms of "AI is building itself". And having a common language —automation levels, oversight coverage, share of compute on safety— is the precondition for any slowdown agreement between labs, or any regulatory transparency obligation, to be verifiable rather than a good-faith promise.

But we must be equally clear about the limits. These are metrics Anthropic has designed, measured and published about itself, with its own model as judge and its own definition of what counts as "safety work" versus "capabilities work" —a boundary that, as the company acknowledges, every lab will be tempted to draw generously—. The company admits as much, which is why it proposes that the burden of proof fall on the developer and that third parties verify the classification. Until that genuinely happens —with external evaluators with real access, not merely announced ones—, this is as much measurement science as it is an institutional relations play: Anthropic positions itself as the lab calling for traffic rules before anyone crashes, which also serves its interests against rivals less willing to show their numbers.

The figure that should give most pause is not the 26% automation, but the 6% of compute devoted to safety against the more than 90% of work that already has some degree of AI involvement. If R&D automation advances faster than proportional investment in understanding and controlling it, that imbalance —not the science fiction of a model rewriting itself overnight— is the real short-term risk. It is exactly the kind of friction that underpins our underlying thesis: the path towards an AI that helps eradicate disease and multiply abundance is not a straight line, and it runs through governing these years of self-acceleration well, not through pretending they do not exist.