Anthropic said its Claude model now leads 26% of the company’s AI research and development work, up from under 1% in February, in a disclosure the company framed as the first installment of regular reporting on how much of its research is done by AI itself.
The figures, published Thursday and carried by Reuters, Bloomberg and the Associated Press, offer one of the most detailed public windows yet into how frontier labs actually build their models. More than 90% of Anthropic’s R&D work now happens at a level where AI either collaborates with humans or leads the work outright. The company is effectively reporting that its next generation of models is being built, in significant part, by the current one.
How the measurement works
Anthropic built what it calls the R&D Automation Index, rating research tasks on an automation scale developed by Epoch AI. The scale runs from AL0, no AI involvement, to AL5, fully autonomous operation with no human in the loop. At AL3, AI collaborates, meaning it handles large chunks of work under close human direction. At AL4, the level Anthropic labels leads, AI can complete most of a task end-to-end from a high-level prompt while a human supervises.
As of August, Claude had reached AL4 on 26% of the company’s AI R&D work. The company stressed that the model has not reached AL5, full autonomy, in any measured area. The distinction matters for how the number should be read: this is supervised delegation at scale, not machines running the lab alone.
The scale of the operation is substantial. Around 30,000 AI agents were carrying out research and engineering work on Anthropic’s most-used internal platform at any given time in August. Every action those agents take passes through an online monitor before execution, usually within seconds, and a separate offline system reviews activity after the fact.
| Measure | Figure |
|---|---|
| R&D led by Claude (AL4) | 26%, up from under 1% in February |
| R&D with AI collaborating or leading | Over 90% |
| AI agents on internal platform | About 30,000 at any time |
| Agent decisions blocked by monitor | 1 in 47,000 (0.002%) |
| Compute share devoted to safety work | 6% of R&D, 12% of AI-run research |
The monitoring numbers
Anthropic analyzed more than 1 billion agent decisions from August. About 0.002%, or one in 47,000, were blocked by the online monitor before execution. The system also flags roughly 100,000 transcripts a week for further review, with about 50 of the highest-priority cases escalated to human reviewers.
Those ratios will be parsed differently depending on priors. A monitor that blocks one action in 47,000 across a billion monthly decisions still catches roughly 21,000 blocked actions a month. Whether that indicates tight control or a wide net depends on what the blocked actions would have done. Anthropic did not publish details of what the blocked actions involved, which leaves the headline figure open to interpretation in both directions.
On compute, the company said about 6% of the computing power used for AI research went to safety work in a sample week in July, rising to 12% for research carried out by AI itself. Anthropic called those figures conservative, because compute that advanced safety and capability equally was counted as capability work. The company did not say what share of safety compute goes to evaluating the agent infrastructure itself versus model-level alignment research.
The context around the disclosure
The announcement lands in a charged week for AI safety debate. Anthropic CEO Dario Amodei has been among the lab leaders calling for a slowdown in development over safety concerns, while President Trump dismissed safety warnings as a hoax, announced an AI Force modeled on Space Force and said he would name an AI czar. OpenAI, meanwhile, said Wednesday it would begin regularly publishing reports on unexpected or unauthorized model behavior, disclosing six such reports of its own.
Researchers have warned for years that as AI agents become more autonomous, they may develop behaviors that diverge from their creators’ intentions and become harder to monitor. The Anthropic numbers put a concrete measurement on how close the industry is getting to that line. Recursive self-improvement, where AI systems meaningfully advance their own capabilities without human input, remains the threshold case. The company’s own framing is that 26% led, zero autonomous is a measurable distance from that threshold, and one it intends to keep reporting on publicly.
“Those in our world who value our humanity and its vital moral component are anxiously seeking your reassurance that we will not lose control of our destiny,” King Charles told executives from Nvidia, OpenAI and Anthropic at a Scotland summit the same week, capturing the external pressure the labs are operating under while they publish internal metrics like these.
The disclosure also arrives amid commercial pressure that cuts against any slowdown. Anthropic has signed a $35 billion cloud deal with Nvidia-backed Lambda, a 20-year, $9.1 billion lease with bitcoin miner Riot Platforms and a $19 billion agreement with TeraWulf, and it reportedly recorded its first profitable quarter in June. The company is reportedly weighing a $6 billion acquisition of AI startup Decart and considering a new model launch to counter OpenAI’s GPT-6 Astra, with its IPO now expected after the November midterms. A lab racing a public listing does not slow down on rhetoric alone, which is why the automation index reads as both a transparency measure and a bet that measured disclosure buys it room to keep accelerating.
What the index does not cover
The index measures Anthropic’s model R&D only. It does not cover product engineering, infrastructure, or the operational work of running a company with a fast-growing commercial business. Those functions are also being automated across the industry, but Anthropic has not published equivalent figures for them. Critics of self-reported metrics will note the company chose the scope, the scale and the timing of its own transparency.
Anthropic invited competitors to publish similar figures. Whether OpenAI or Google DeepMind adopt comparable reporting will determine whether the R&D Automation Index becomes a sector standard or a one-company disclosure. OpenAI’s new behavior-reporting commitment is a start, but it covers unexpected model behavior rather than measuring how much research AI does.
For now, Anthropic has set the baseline: a quarter of its model research led by its own model, 30,000 agents on the internal platform, monitored action by action, with the next report expected to show whether that share keeps climbing. If February’s near-zero to August’s 26% trajectory continues at anything like the same pace, the next disclosure will be the one to watch.