Full transcript
The narration below matches the explainer. Its original reporting dates and qualifications still apply.
ANTHROPIC’S BENCHMARK PROGRESS SPED UP

METR estimates that Anthropic’s AI benchmark scores improved one point nine times as fast after October twenty twenty-five, compared with the earlier trend.
METR assessment · 2026-11-05
THE SPEEDUP ESTIMATE HAS A WIDE RANGE

The confidence interval runs from one point five to two point eight times the earlier pace. That range describes uncertainty in the benchmark estimate.
METR assessment · 2026-11-05
METR POINTS TO AI ASSISTANCE

Staff interviews and surveys suggest that AI assistance sped up research. METR thinks it likely explains most of the recent acceleration, with substantial uncertainty about its share.
METR assessment · 2026-11-05
THE PUBLIC SAW ONLY PART OF THE EVIDENCE

METR had access to internal benchmarks and researchers. Anthropic redacted most of the evidence METR considered most informative, limiting what outside readers can assess.
METR assessment · 2026-11-05
A FASTER BENCHMARK TREND LEAVES OPEN QUESTIONS

The benchmark results point to faster improvement on measured tasks. How much AI contributed and how far the finding generalizes remain open questions.
METR assessment · 2026-11-05
SOURCES AND QUALIFICATIONS
METR published the assessment and its statistical range. The Associated Press reported on the findings and Anthropic’s response.
METR assessment · 2026-11-05