What did the outside review find?
The Associated Press reports that independent research group METR estimates Anthropic’s progress on well-specified AI tasks is running about twice as fast as its earlier pace. The comparison spans the earlier March 2024 to October 2025 period and the newer review window. This is an estimate of the pace of improvement on measured tasks, not a finding that every AI ability doubles each year.
How certain is the estimate?
METR’s own assessment calls the result highly uncertain: a range nearer 1.3 to 3 times the earlier pace would not surprise its researchers. Its benchmarks emphasize tasks with clear success criteria. Anthropic gave METR unusual access to internal records, but redacted much of the detail most relevant to its conclusions, limiting outside verification.
Is AI improving itself?
METR thinks AI assistance likely contributes to faster research, but cannot assign a precise causal share. Human researchers still direct the work; the report does not establish autonomous self-improvement or a runaway system. Other changes in staffing, methods and investment may also matter.
Why does this matter for safety?
Anthropic CEO Dario Amodei has urged the wider industry to slow improvement while safety work catches up. That call is distinct from a confirmed pause inside Anthropic. The practical question is whether independent evaluations and safeguards can keep up with the measured pace. Watch for future assessments with more publishable evidence and broader tasks. Listen to the full episode or watch the Short.
