Anthropic says Claude now leads 26% of its own AI research work — up from under 1% seven months ago
Anthropic published its first R&D Automation Index on 17 September 2026, showing Claude went from leading under 1% of the company's AI research and development tasks in February to 26% by August, with more than 90% of R&D work now involving Claude in some capacity — a concrete curve for how fast AI is automating the work of building AI.
19 September 2026
Anthropic built something unusual for a lab this size: a metric that tracks how much of its own engineering work AI is actually doing, and published the trend line. The R&D Automation Index samples about 20% of staff across departments involved in model research every week, logging roughly 15,000 individual tasks a month and rating each one on a five-point automation scale (developed with Epoch AI) from AL0 — no AI involvement — to AL5, full autonomy with no human in the loop. Claude was rated AL4 (“AI leads”: the model completes most of a task end-to-end from a high-level prompt, with a human supervising) or higher on 26% of Anthropic’s AI R&D tasks as of August 2026. In February, that figure was under 1%. It moved to 12% by May and 22% by July — a curve that’s still accelerating, not levelling off.
The scale of deployment behind that number is the part worth pausing on: about 30,000 Claude agents were running at once on Anthropic’s internal platform in August, and Claude now touches more than 90% of the company’s R&D work in some capacity, even where it isn’t yet leading. Anthropic also reports that its automated safety monitor blocked roughly one in every 47,000 of over a billion agent decisions — evidence the company is building oversight infrastructure at the same pace as the automation itself, not bolting it on afterward.
Why a self-reported number is still useful
Anthropic has an obvious interest in this story looking good, and the index is built from the company’s own Slack messages and internal documentation rather than an independent audit. Take the exact percentage with that caveat in mind. But the shape of the curve — near-zero to over a quarter of R&D work in seven months — is consistent with what other labs and enterprise engineering teams have been reporting independently this year: AI-authored code volume, CI load and agent task delegation are all climbing on similar timelines elsewhere, not just inside Anthropic.
So what
For anyone commissioning software, the practical takeaway isn’t “AI now does a quarter of the work” — it’s that the rate of change is the thing to plan around, not any single snapshot. A development partner’s tooling and workflow six months ago is not a reliable guide to what they’re capable of today, and a team that hasn’t measurably changed how it works with AI agents in that time is worth asking about directly. We treat this shift as core to how we scope and deliver projects — see our AI-assisted development approach, or get in touch if you want a straight answer on how much of your build would realistically involve AI-led work versus human-led engineering.