METR Report Suggests Anthropic Has Crossed 'Dramatic Acceleration' Threshold

Details

A METR report indicates that between October 2025 and October 2026, Anthropic's capability progress (measured by the internal Epoch Capabilities Index, AECI) was approximately 2x faster than the previous 12 months.
The report attributes roughly 80% of the current speed-up relative to a no-AI counterfactual to AI automation, implying an observed speed increase of ~1.8x relative to the counterfactual.
The author of the analysis argues that these findings meet the 'dramatic acceleration' threshold defined in Anthropic's Responsible Scaling Policy (RSP), which requires progress to be at least 2x faster than the fastest sustainable pre-AI rate and attributable to AI automation.
The report authors did not identify clear bottlenecks or steep diminishing returns that would prevent compounding acceleration, though they noted significant uncertainty and evolving understanding.
The author notes that current industry security levels are insufficient for the projected timeline, citing recent breaches at OpenAI and Anthropic by amateur groups, which fall short of the ASL-4 standard required for protecting model weights against state actors.
The author's heuristic modeling suggests that maintaining the current acceleration may require a ~10x labor uplift for researchers, compared to an initial theoretical estimate of 4–9x, due to the exhaustion of low-hanging fruit in idea discovery.

Key point