Recursive self-improvement just got a public scoreboard. On Thursday, Anthropic said its Claude systems now “lead” 26% of the company’s internal AI research and development—meaning they can finish most of a task end-to-end from a high-level prompt while a human supervises—up from roughly 1% in March on an Epoch AI automation scale.
Reuters reports the San Francisco lab also said AI collaborated with humans on more than 90% of research work as of August. Claude is still not fully autonomous in any measured slice of that work. That caveat matters, but the trajectory is the story: in half a year, a frontier lab went from “AI occasionally helps” to “AI leads a quarter of how we build the next model.”
What “leads” actually means
Anthropic’s blog lays out an Anthropic R&D Automation Index built from a frozen tree of ~542 task categories distilled from roughly 15,000 granular R&D tasks sampled from staff work records. Ratings use Epoch AI’s Automation Levels: AL3 is “collaborates,” AL4 is “leads,” AL5 is fully autonomous. As of August 2026: no AL5 anywhere they measured; 26% at AL4; over 90% at AL3 or higher.
The company is publishing this as one of three recurring transparency metrics—alongside agent oversight stats and how much R&D compute goes to safety—explicitly so outsiders can watch whether AI is accelerating AI faster than humans (and regulators) can keep up. CEO Dario Amodei has been pushing industry pacing; these numbers are the receipts.
30,000 agents and a one-in-47,000 block rate
On Anthropic’s main internal platform in August, about 30,000 AI agents were doing research and engineering at any moment. Every action supposedly hits an online monitor before execution; over a billion decisions that month, roughly 1 in 47,000 (~0.002%) was blocked. Offline monitors flag roughly one to two transcripts per thousand for further review, with ~50 high-priority flags escalated to humans each week.
That’s the control story Anthropic wants next to the acceleration story. The lab also said that in a sample week in July, about 6% of AI R&D compute went to safety work—rising to 12% for compute used by AI-driven R&D—and that those figures are deliberately conservative (dual-purpose work counted as capability).
Timing: OpenAI’s misbehavior disclosures
The drop lands a day after rival OpenAI said it would regularly publish unexpected or unauthorized model behavior and disclosed six concerning incidents. Anthropic’s message is different but rhymes: if labs are going to argue for pacing and trust, they need public dashboards—not vibes—about how much of the sausage factory is already automated.
Cross-lab comparison is still hard. Anthropic admits methodology isn’t standardized and that using Claude to help rate Claude’s own automation invites judge-model bias. Their proposed fix: third-party evaluators embedded with access comparable to internal risk teams. Until others publish the same AL breakdowns, “26%” is a single-lab lighthouse, not a league table.
Geeknewz take
This is the most concrete public peek yet at how close a frontier lab is to AI that builds AI. Twenty-six percent “leads” with zero full autonomy is not Skynet—it’s a power tool with a human still holding the kill switch. The scary part isn’t the absolute number; it’s the slope from 1% to 26% in months while ~30k agents thrash through R&D under automated monitors.
Watch three things: whether that 26% keeps climbing into the 40s without hitting AL5; whether other labs publish comparable indices (or refuse); and whether the 6%/12% safety-compute shares rise if Amodei’s pacing talk turns into actual coordination. Transparency theater is still useful if the numbers are hard to fake. Anthropic just raised the bar for everyone else to put their automation on a chart.
Source: Reuters — Anthropic says Claude now leads a quarter of work building its next AI models (Aditya Soni, Sep 17, 2026); also Anthropic Institute blog.