The findings
METR’s November 5 report estimates 1.9× acceleration in Anthropic’s AECI trend since October 2025, with a 95% confidence interval of 1.5–2.8×. Its separate subjective assessment of progress on well-specified tasks is roughly 2×, with 1.3–3× plausible.
Anthropic redacted most evidence METR considered most informative. The report does not establish fully autonomous AI self-improvement.
What an index measures
Epoch explains that ECI combines different benchmark results into a common scale, allowing model comparisons even after individual tests saturate. An index value is abstract on its own. It can be related to expected performance on particular benchmarks.
Why context matters
A historical example: METR’s July 2025 study involved 16 experienced developers completing 246 tasks. Allowing the AI tools of that period increased completion time by 19%. This was a particular experiment with early-2025 tools, not a measurement of current models.
What teams should measure
Our practical interpretation: before putting an acceleration assumption into a budget or deadline, define your own unit of output. For engineering, that might be an accepted change with tests; for research, a reproduced experiment. Include review, corrections and failed attempts, rather than measuring only the time to a first answer.
Track quality, spending and completed work separately. If a tool produces more code but lengthens the review queue, the overall benefit still needs testing. Compare similar tasks and record the model version and usage rules.
Safety discussions also need a separate question: can the outputs be checked and the process stopped before an error spreads? Productivity growth and effective oversight require different measures.
Comments