Anthropic says Claude now leads 26% of its own R&D — the first public scorecard for AI building AI
Claude went from roughly 1% of Anthropic's model R&D to a quarter in a matter of months. Anthropic is publishing its internal scorecard — and asking every rival lab to do the same.
On Thursday, Anthropic published a number no frontier lab had ever disclosed: the share of its own AI research that is now led by its own AI. As of August, Claude led 26 percent of the company’s model research and development — carrying research tasks most of the way from a high-level prompt to completion, still under human supervision. Earlier this year, the figure was roughly 1 percent.
The disclosure is the first serious attempt to measure something the industry has discussed in private for years: recursive self-improvement, the point where models help engineer their successors. Anthropic is not just publishing its own score. It is asking every rival lab to publish the same kind of numbers, on a shared public methodology, so the whole field can watch the curve in real time.
The numbers, and what “leading” means#
“Leading” means doing most of the work. Anthropic’s definition of the top tier: Claude completes most of a task end to end from a high-level instruction, with a human supervisor reviewing the result rather than steering each step. The model is not working completely on its own — the company stresses that nothing in its pipeline is yet fully autonomous.
Below that sits “collaborating”: Claude doing large chunks of work under close human direction. More than 90 percent of Anthropic’s R&D now happens at or above that level. The classification comes from an automation scale built by Epoch AI, an independent nonprofit that tracks AI progress — a deliberate choice, since it gives other labs a yardstick they could actually adopt.
The scale of the operation is concrete. Roughly 30,000 Claude agents run simultaneously on Anthropic’s internal research platform, doing work that looks like junior-researcher work: diagnosing why a training run kept failing, why a model underperformed on a benchmark, how to restructure a piece of infrastructure that had grown unwieldy.
| Metric | Figure |
|---|---|
| R&D share led by Claude (August) | 26%, up from ~1% earlier this year |
| R&D with Claude collaborating or leading | Over 90% |
| Internal agents running at once | ~30,000 |
| Agent decisions in August | Over 1 billion |
| Actions blocked before execution | 0.002% (~1 in 47,000) |
| R&D compute spent on agent safety (one July week) | 6% |
The safety budget is real money#
The most revealing part of the disclosure may be the oversight ledger. With a billion agent decisions a month, Anthropic runs a layered monitoring system: a real-time layer that blocked 0.002 percent of actions before they executed, and an offline layer that flagged about 100,000 transcripts a week for review — of which roughly 50 were escalated to a human.
During a single week in July, work dedicated to keeping agents safe consumed six percent of all AI R&D compute — and twelve percent of the compute used by the AI-driven R&D itself. Anthropic says it counted conservatively, excluding dedicated safety classifiers, which suggests the true share runs higher.
This turns a philosophical debate into an accounting one. As AI-led research grows, the safety overhead grows with it — and the ratio between the two becomes one of the few visible indicators of whether a lab’s caution is scaling with its ambition.
Why publish it now#
Timing is part of the story. The disclosure landed days after Anthropic CEO Dario Amodei’s essay calling on the industry to slow the pace of frontier capability gains — a call that rattled AI stocks last week. Against that backdrop, publishing the numbers reads as an attempt to make the debate concrete: here, measured and auditable, is how fast AI is taking over the work of improving AI.
The company framed the release in those terms. In a blog post, Anthropic argued for shrinking “the gap between what frontier labs know and what the public knows” — measuring AI development, reporting it publicly, and letting society decide what to do with the information. It warned that models accelerating their own development could become harder for humans to understand or control, and argued that shared metrics could show how close the field is getting to genuine recursive self-improvement.
Whether rivals comply is an open question. No other lab publishes an equivalent figure, and there is little obvious incentive to be the second one. But the call itself changes the game: any lab that stays silent now does so against a public benchmark it did not set.
What to watch#
First, the next measurement. If the curve keeps rising at this slope, the “leads” share doubles again within a year — the point at which a majority of the work behind a new model was itself done by a model.
Second, whether a shared methodology materializes. Anthropic is using Epoch AI’s scale, which helps, but cross-lab comparison needs more than one volunteer. Watch for OpenAI, Google DeepMind, and xAI to respond — with numbers, or with excuses.
Third, the oversight ratio. Six percent of R&D compute going to agent safety sounds small until you remember it was effectively zero not long ago. If AI-led research doubles again while the safety share stays flat, someone will eventually have to explain the gap — publicly, now that the numbers are out there.