What changed
METR, a nonprofit focused on AI safety, found that Anthropic's models are improving at roughly double the previous rate.
The analysis, based on internal data from August to October 2026, suggests that about 80 percent of this acceleration stems from AI being used in Anthropic's own research.
The nonprofit's findings indicate that the pace of AI progress has doubled since October 2025, with the AECI metric showing a jump from 15 to 29 points per year.
What this means in practice
The acceleration is attributed to AI acting as an assistant in research, reducing bottlenecks in code and setup. Employees reported that experiments now take about 12 times less time to set up, though METR treats these self-assessments as loose evidence.
The nonprofit also notes that AI is not yet doing fully autonomous research, but there are signs it could play a larger role in the future.
METR warns that the current pace of acceleration could continue, increasing the risk of sudden emergence of extreme capabilities.
Limitations and availability
METR acknowledges significant uncertainty in the results, with the true acceleration factor possibly ranging from 1.3 to 3. The analysis also highlights that Anthropic redacted much of the evidence, including forecasts and bottleneck analysis, which weakens the audit's transparency.
The nonprofit's findings suggest that the warning threshold may have been crossed as early as July 15, but the public report from Anthropic understates the acceleration and its uncertainty.
The audit also found no immediate bottlenecks that would slow the pace over the next year.
