A high-access evaluation study conducted by the non-profit organization METR concluded that Anthropic has doubled the pace of improving its models compared to last year, driven by the use of artificial intelligence in internal research.
METR’s independent assessment and the doubled pace of Anthropic’s model development
The role of artificial intelligence in accelerating research by 80% and methodological caveats
The contrast with Anthropic’s internal risk report and surpassing the double acceleration threshold
Dario Amodei’s essay calling for slowing the pace and preparations for the public offering
An analytical reading of the strategic and economic dimensions of the event
Frequently asked questions
METR’s independent assessment and the doubled pace of Anthropic’s model development
Anthropic is improving its artificial intelligence model capabilities at a speed nearly double the pace it was moving at a year ago, according to an independent scientific evaluation released on Thursday by the non-profit model evaluation organization METR. This striking finding comes just weeks after company CEO Dario Amodei called on the global artificial intelligence industry to slow the pace of its progress, and the report represents one of the most detailed external studies to date on whether artificial intelligence is accelerating its own development process, coinciding with Anthropic’s expected initial public offering later this month, according to the Associated Press.
METR researchers conducted a “high-access evaluation” inside Anthropic between August 12 and October 29; the report concluded that as of October, the company “is likely improving its model capabilities on well-defined tasks at a rate approximately twice the rate” recorded between March 2024 and October 2025. The group measured this progress using Anthropic’s internal benchmark index and found that results grew about 1.9 times faster since October 2025, with a 95% statistical confidence interval ranging between 1.5 and 2.8 times, meaning the company achieved roughly two years of its previous progress in the past year alone.
The role of artificial intelligence in accelerating research by 80% and methodological caveats
METR stated that it is “confident” that faster growth in the amount of compute dedicated to training was not the cause of this acceleration; it explained that most of the current acceleration likely stems from Anthropic’s heavy use of artificial intelligence tools within its own research. Reuters reported that METR estimated artificial intelligence’s share of this acceleration at roughly 80%, while describing this figure as not entirely certain, and researchers found no convincing evidence of major bottlenecks that could slow progress over the coming year.
However, the report included important methodological caveats; METR noted that the true rate might be closer to 1.3 times or 3 times the historical pace, that the acceleration may apply only to tasks that can be automatically evaluated and corrected, and that editorial redactions imposed by Anthropic to protect its secrets limited the amount of detail the group could publish, according to the Associated Press.
The contrast with Anthropic’s internal risk report and surpassing the double acceleration threshold
In a separate review, METR said that Anthropic’s risk report published last August “underestimates the level of acceleration that was already taking place.”
The organization explained that Anthropic’s internal “double acceleration threshold” had been reasonably surpassed by July 15. Nonetheless, METR agreed with the company that the immediate catastrophic risks posed by those models were low, but added that it “cannot rule out a significant additional acceleration” over the next six months.
Dario Amodei’s essay calling for slowing the pace and preparations for the public offering
CEO Dario Amodei had urged the global artificial intelligence community in a September 12 essay to “slow the pace at which we improve artificial intelligence model capabilities,” warning that repeated self-improvement “could outpace our ability to understand and control these systems.” In the same essay, he named METR as an appropriate external evaluator, according to the Associated Press, and the company also published its own research showing that artificial intelligence is already accelerating its development work.
Anthropic did not immediately respond to requests for comment on the report; METR stated that artificial intelligence cannot yet provide “fully independent, highly critical contributions” to model training, but noted that Anthropic researchers described “early life signs” suggesting that could change soon.
An analytical reading of the strategic and economic dimensions of the event
The results of the independent evaluation issued by the non-profit METR organization acquire extreme scientific and regulatory importance because they provide the first external “high-access” measurement of the phenomenon of artificial intelligence development acceleration by artificial intelligence itself within one of the world’s most prominent labs. Concluding that Anthropic is improving its model capabilities on well-defined tasks at a rate about 1.9 times faster since October 2025—achieving two years of progress in just one year—while attributing about 80% of this acceleration to the use of artificial intelligence in internal research rather than increased training compute, lends realistic weight to the ongoing debate over repeated self-improvement of models.
The contrast between METR’s estimates and Anthropic’s internal risk report published in August also highlights the importance of independent auditing; the organization suggested that the double acceleration threshold had already been crossed by mid-July, without ruling out significant additional acceleration over the next six months. Although both sides agree that immediate catastrophic risks are currently low and that artificial intelligence has not yet reached fully independent self-training, these factors place the warnings of CEO Dario Amodei calling for slowing the pace of the race at the forefront of investors’ and regulators’ concerns ahead of the company’s initial public offering.
Frequently asked questions
Question: What did the METR study conclude regarding the speed of Anthropic’s model development?
Answer: The study, conducted between August 12 and October 29, concluded that Anthropic is developing its model capabilities on well-defined tasks at a rate about 1.9 times faster (nearly double) compared to the period between March 2024 and October 2025.
Question: What is the main reason behind this acceleration according to the METR report?
Answer: The organization ruled out increased training compute as the cause, attributing about 80% of the acceleration to Anthropic’s intensive use of artificial intelligence in its internal research and development.
Question: How did METR criticize Anthropic’s August risk report?
Answer: It stated that the August report underestimated the level of actual acceleration, suggesting that the company had crossed its internal double acceleration threshold by July 15, while agreeing that immediate catastrophic risks remain low.
Question: What was CEO Dario Amodei’s stance on accelerating artificial intelligence capabilities?
Answer: Amodei urged in a September 12 essay to slow the pace of developing model capabilities for fear that repeated self-improvement would outpace human ability to understand and control it, and proposed METR as an external evaluator.