The finding comes from an eleven-week, high-access review in which METR staff examined internal Anthropic data on how much AI is speeding up AI research and development.
Using Anthropic's version of the Epoch Capabilities Index, a benchmark-based measure of model capability, METR put the acceleration at 1.9x, with a 95% confidence interval of 1.5 to 2.8x.
That would mean roughly two years of earlier progress happened in the past year.
METR calls the estimate highly uncertain and says the true figure could be nearer 1.3x or 3x.
Its staff estimate that about 80% of the current acceleration comes from AI, reflecting Anthropic's heavy use of AI in its own research, and they are confident it was not caused by faster growth in training compute.
They found no convincing bottlenecks in algorithms, compute, or data over the next 12 months, and say they cannot rule out substantial further acceleration over the next six months.
In a separate review of Anthropic's public August 2026 Risk Report, METR agreed that catastrophic risks from models as of July 15 were low.
It argued, however, that the report's claim that the 2x acceleration early-warning threshold had not been crossed is overstated.
The private report's figures imply 1.97, which METR says is better described as "around 2" than "less than a factor of 2."
Transparency is a major caveat. Anthropic had the right to redact non-public information and removed most of the evidence METR found most informative, saying the cuts targeted confidential material that could help other AI developers. METR says the public must therefore lean on its judgment more than is ideal. Anthropic also dropped a planned check of how its computing power is divided between capabilities and safety work, saying it would revisit the topic later. METR notes further limits: benchmark scores may not reflect real-world impact, Anthropic's benchmarks are increasingly saturated, and estimates ranged from 1.1x to 2.9x depending on analysis choices. METR says it received no payment from Anthropic, though Anthropic supplied laptops and free tokens, and it disclosed a conflict of interest involving one team member.