The AI safety research group METR has published a report, released on 5 November, on how much AI is speeding up research and development at Anthropic. After 11 weeks of access inside the company, it concludes that Anthropic's models are now improving roughly twice as fast as they were a year earlier, and that most of that speed-up comes from AI itself rather than from more computing power.
The findings have alarmed safety researchers, both for what they say about how fast AI can now improve itself and for what the public version leaves out: Anthropic used its right to redact the report, and by METR's account it removed most of the evidence the researchers considered most informative.
What METR found
According to The Verge, METR's benchmark data suggest Anthropic made about two years' worth of its historical progress in the past year. The researchers say they are confident the speed-up is not just a matter of training with more compute, and that by October the majority of it was due to AI, with that share rising over time. Staff using the company's strongest internal model reported, on average, being about 12 times as productive as without AI help, around double what they reported a few months earlier.
Most of the gain still comes from AI assisting human researchers rather than working alone, though METR sees signs that fully autonomous work will soon matter more. It found no clear sign of diminishing returns or bottlenecks that would stop the pace from rising further. The report also says Anthropic finished training three models in the past six months that it has not released, completing in July, August to September, and September to October.
Business Insider adds context from Anthropic's IPO filing. Valued at the prices Anthropic charges customers, the company's internal use of Claude passed its monthly employee pay, excluding stock awards, in April, and internal usage grew more than 250-fold between January and September. Claude also wrote more than 90 percent of the new code that went into Anthropic's own software that September.
Why researchers are worried
In an essay on planned-obsolescence.org, a METR staff member who did not work on the project argues that the findings mean Anthropic has already crossed the threshold its own Responsible Scaling Policy calls dramatically accelerated progress due to AI. The author notes that the METR team credits roughly 80 percent of the current speed-up to AI, which works out to about 1.8 times the rate without AI as of October, and expects that figure to have grown since.
The essay's simple model goes further: if acceleration keeps compounding smoothly, each doubling of the rate of progress would arrive faster than the last. The author stresses that this naive extrapolation is very uncertain, and argues that strong protection of model weights and research secrets should be a bare minimum before such a phase, a bar the author says the industry is far from meeting.
Other researchers quoted by The Verge reached similar conclusions. Marius Hobbhahn of Apollo Research said the pace looks faster than the influential AI 2027 scenario, and METR researcher Ryan Greenblatt wrote on X that he had expected much less speed-up in 2026. Alan Chan of GovAI said skeptics may need to shorten their estimates of when tasks such as software engineering are fully automated.
The fight over redactions
METR said it gave Anthropic a report of about 100 pages, that a few of its data requests were refused, and that the redactions leave the public relying on METR's judgment far more than it should. It described most of the removed material as very high-level. Anthropic co-founder Sam McCandlish replied on X that the cuts make METR's work harder to check, but that revealing what drives the speed-up is commercially sensitive and could give rivals a roadmap to go faster. He said Anthropic will publish its own report on the pace of AI progress in the coming weeks.
Critics see a contradiction in a company that has urged others to slow down. Daniel Kokotajlo of the AI Futures Project accused it of gambling recklessly, Hobbhahn pointed to Anthropic cancelling the part of the review meant to check how much compute goes to safety work, and New York Assemblyman Alex Bores called on Congress to ban recursive self-improvement. Anthropic declined to answer The Verge's questions beyond pointing to McCandlish's post.
Sources
- Why a new report on Anthropic automating AI development is alarming researchers The Verge
- AI is dramatically accelerating AI research planned-obsolescence.org
- Anthropic's internal AI use, at retail prices, now exceeds its employee pay businessinsider.com