Anthropic Says Claude Now Leads 26% of the Work That Builds the Next Claude

Anthropic published a report on September 17, 2026, proposing three measurements the public could use to track how fast AI development is moving inside frontier labs: how much AI research is done by AI itself, how closely AI agents are overseen, and how much computing power goes toward safety. The company published a snapshot of all three from inside its own walls, and the headline figure is that Claude now “leads” more than a quarter of the work that goes into building the next Claude.

What Exactly Did Anthropic Measure?

Anthropic frames the release around a concern it has raised before: AI systems are increasingly being used to build the next version of themselves, and the pace of that self-acceleration is hard for anyone outside a frontier lab to see. The company says it wants to narrow the gap between what labs know internally and what the public, governments, and third parties know. It ties the release directly to Anthropic CEO Dario Amodei’s public calls for coordination on pacing frontier AI development, noting that these numbers would be expected to shift if such coordination happened.

The report covers three separate measurements, each with its own methodology, findings, and a description of what any other AI developer could publish using the same approach.

How Much of Claude’s Own Development Does Claude Now Handle?

To track this, Anthropic built what it calls the Anthropic R&D Automation Index. Using an automation scale developed by Epoch AI, every piece of AI R&D work gets rated from AL0 (no AI involvement) up to AL5 (AI operates fully autonomously, with no human in the loop). AL3 means AI “collaborates,” handling large chunks of work under close human direction; AL4 means AI “leads,” completing most of a task end-to-end from a high-level prompt while a human supervises the result. This mirrors the shift already visible in products like Claude Code Projects, where a single instruction can now fan out into work a human only checks in on periodically rather than directs step by step.

To build the index, Anthropic sampled 20% of staff across departments in the model R&D loop each week through July 2026, had a Claude research agent catalog what each person worked on using Slack and internal documentation, and organized roughly 15,000 granular tasks into a frozen tree of 542 categories. A separate Claude judge model then rated the automation level of each category, weighted by how much staff time it represented.

As of August 2026, Anthropic reports:

• Claude is not operating fully autonomously (AL5) on any measured subset of AI R&D work.
• Claude “leads” (AL4) 26% of Anthropic’s measured AI R&D work, up from under 1% in February 2026.
• The share of work at or above “AI collaborates” (AL3 and up) is above 90%.

Anthropic acknowledges the obvious limitation: it is using its own models to judge its own systems, which risks the judge repeating the same blind spots as the model it is grading. To check this, staff who own each work area rated their own team’s automation level without seeing what evidence the models had gathered. The model and human ratings landed within one level of each other 97% of the time, though exact agreement between model and human was 59%, versus 35% between two different humans rating the same work.

How Closely Is Anthropic Watching Its AI Agents?

The second measurement covers oversight: as work shifts from “AI collaborates” toward “AI leads,” AI agents start making more consequential decisions on their own, including which research direction to pursue next. Anthropic says roughly 30,000 agents were doing research and engineering work at the company at any one time in August 2026, on its most-used internal platform.

Those agents run under two kinds of monitors. Online monitors check actions in real time and can block or redirect anything dangerous and irreversible, like an agent attempting to copy model weights out of Anthropic’s systems — 100% of agent actions pass through this layer before execution, usually within seconds. Offline monitors read agent activity after the fact, looking for slower-developing problems like signs of misalignment; they flag roughly 100,000 transcripts a week for review, with the highest-priority cases (about 50 a week) escalated to a human.

Anthropic analyzed more than a billion decisions made by its research and engineering agents in August 2026 and found that 0.002% — about 1 in every 47,000 — were blocked by the online monitor. The offline monitor flags roughly one to two transcripts per thousand for further review. Anthropic says individual agents rarely misbehave in its data so far, but with millions or billions of agents eventually operating across the economy, even rare failure rates would start happening regularly.

Where Is Safety Compute Actually Going?

The third measurement looks at how Anthropic allocates computing power. Examining a one-week snapshot from July 13 to July 20, 2026, Anthropic found that about 6% of the compute spent on AI R&D went toward safety work, rising to 12% when measured specifically against compute used for AI-driven AI R&D. Anthropic calls these conservative estimates: if a given token advanced both capabilities and safety, it wasn’t counted, and the figures exclude a separate, comparable amount of compute spent on safeguard classifiers.

Why Release This Now?

Anthropic frames the release as a template, not a one-off: it says any frontier developer could publish the same three measurements — automation share, monitoring coverage and escalation rate, and safety compute share — on a regular basis using a public methodology, which would make the numbers comparable over time and potentially across labs. The company says it plans to embed independent third-party evaluators inside Anthropic with access to internal processes, systems, and data comparable to its own risk assessment teams, to verify safety practices and monitor these same metrics going forward. It notes that the external research group METR has previously red-teamed its offline monitoring platform independently.

Why This Matters

The framing worth holding onto is Anthropic’s own: Claude is not autonomously building Claude. Twenty-six percent “leads” means a human still sets the goal and supervises the outcome on that slice of work; nothing in the report reached full autonomy. That distinction is easy to lose in a headline number, and it’s the difference between a real, fast-moving automation trend and a much more dramatic claim Anthropic isn’t making.

What is new here isn’t just the number, it’s that Anthropic is proposing the dashboard before anyone has agreed on a speed limit. A company defining the metrics that might eventually justify slowing itself down, or that regulators might use to decide when to step in, is also positioning itself as the one setting the terms of that debate. Anthropic seems aware of this: it explicitly frames third-party verification, not its own self-reporting, as the thing that would make these numbers trustworthy over time. It follows a similar instinct to OpenAI’s own move toward disclosing model misbehavior earlier this month — frontier labs increasingly getting ahead of scrutiny by publishing the uncomfortable numbers themselves. Whether other labs adopt anything close to the same methodology, and whether outside evaluators actually get the access Anthropic says it plans to grant them, is what will determine whether this becomes an industry standard or stays a one-company disclosure.

Frequently Asked Questions

Does this mean Claude is building itself with no human involved?

No. Anthropic is explicit that Claude is not operating fully autonomously (AL5) on any measured subset of AI R&D work. The 26% figure refers to AL4, “leads,” where Claude completes most of a task end-to-end from a high-level human prompt while a person supervises the result.

How did Anthropic measure how automated its AI research is?

It built an R&D Automation Index by cataloguing roughly 15,000 granular tasks performed across its model R&D departments in July 2026, organizing them into 542 categories, and having a Claude judge model rate each category’s automation level on a 0–5 scale adapted from Epoch AI, weighted by how much staff time each category represents.

How many AI agents does Anthropic run internally, and how are they monitored?

Approximately 30,000 agents were doing research and engineering work at Anthropic at any one time as of August 2026. All of their actions pass through an online monitor before execution and an offline monitor after the fact; about 1 in 47,000 actions was blocked by the online monitor over a billion-decision sample from August.

How much of Anthropic’s computing power goes toward AI safety work?

In a one-week snapshot from mid-July 2026, about 6% of compute allocated to AI R&D went to safety work, rising to 12% when measured against compute used specifically for AI-driven AI R&D. Anthropic describes these as conservative estimates that exclude a separate pool of compute spent on safeguard classifiers.

Is this data independently verified?

Not yet. Anthropic acknowledges the figures are self-reported and self-judged using its own models. The company says it plans to embed independent third-party evaluators inside Anthropic with access comparable to its internal risk teams to verify safety practices and monitor these metrics going forward.

Sources: Anthropic, Engadget, The Neuron.

About the author

Kai Williams

Kai Williams has been in marketing for years, with a long background in SEO before AEO had a name. He stepped into Answer Engine Optimization the moment AI started reshaping how people search, and has been tracking the shift ever since. At Prompt Insider, he covers AEO, AI marketing, and the future of search, breaking down what is changing and what brands need to do about it.