METR estimates faster Anthropic capability growth with uncertain AI causation

METR researchers worked within Anthropic from August to October 2026, comparing internal model capability growth against an earlier baseline. Their internal AECI time series suggested accelerated progress, while staff reported extensive use of AI assistants in development.

METR estimated roughly 1.9 times the prior AECI growth rate, with a broad uncertainty interval. Its estimate that AI explains perhaps 80% of the acceleration is a subjective attribution rather than isolated causal measurement, and part of the most informative evidence was redacted.

Three distinct measurements are often compressed into a single claim that AI doubled AI research. METR estimated the rate of Anthropic capability gains at roughly twice its earlier trend, using an internal version of the Epoch AI capability index. The observed AECI growth-rate comparison was about 1.9 times its historical value, with a 95% confidence interval of 1.5 to 2.8. Separately, METR formed a subjective causal view that AI assistance explained most of the acceleration, sometimes summarized as 80%, using staff accounts and internal evidence rather than a randomized counterfactual. Anthropic exercised redaction rights over much of the evidence METR found informative, making independent replication unusually difficult. The report explicitly says it could not exclude important unobserved bottlenecks. Neither 1.9-fold capability-trend growth nor employee self-estimated speedup is proof of autonomous recursive model improvement.

Three quantities should remain separate: observed acceleration in a benchmark-based capability index, staff reports of faster research work, and the portion of company-wide progress caused by AI. METR's approximate 1.9-fold AECI trend is an empirical estimate with an interval; the suggestion that AI explains perhaps 80% of current acceleration is a subjective causal judgment, not the output of a randomized intervention. Internal redactions limit external scrutiny of several decisive inputs. The absence of a visible bottleneck over the next year is therefore a risk assessment, not a guarantee of continued exponential growth or fully autonomous research.