METR’s assessment, published November 5, 2026, estimated that Anthropic’s AECI benchmark scores were growing at 1.9× their earlier rate on well-specified tasks. The 95% confidence interval was 1.5×–2.8×. That is a measure of benchmark-score growth, not a claim that every model capability improved at the same rate. METR’s assessment compares the trend through October 2026 with the earlier period from March 2024 to October 2025.
Read the 1.9× figure as a change in the benchmark trend, not a universal scorecard for every Anthropic model or real-world task.
What METR’s 1.9× estimate measures
AECI, Anthropic’s version of the Epoch Capabilities Index, aggregates scores across a mix of internal and external benchmarks. METR fitted piecewise linear trends to the scores, with October 7, 2025, as the main breakpoint. Its estimate means that the benchmark scores’ growth rate after that point was 1.9 times the earlier rate. METR described the result as roughly two years of historical benchmark progress in one year.
The tasks were well specified, often with results that could be graded automatically or by models. That makes the estimate useful for tracking this benchmark trend; it does not by itself measure how much more useful Anthropic’s models became in everyday settings.
The confidence interval and sensitivity range answer different questions
METR gave the main 1.9× estimate a 95% confidence interval of 1.5×–2.8×. It also tested alternative assumptions about how benchmark scores relate to capability. Under different benchmark link functions, the estimated acceleration ranged from 1.1× to 2.9×.
The first range describes statistical uncertainty around the main estimate. The second shows how the estimate changes when METR varies that modeling choice. They are not interchangeable. METR also noted that its confidence interval does not include sampling uncertainty about which models were included in the analysis.
METR’s estimate of AI’s role is uncertain
METR’s judgment was that AI likely accounted for a majority of the acceleration in October 2026, perhaps about 80%. It described that attribution as subjective and substantially uncertain. METR also judged that non-AI factors likely contributed to the acceleration in late 2025, and said faster growth in training-compute scale did not explain the result.
A separate October 2026 Slack poll put staff members’ average self-estimated AI-related uplift for the most capable internal model at about 12×. METR described those responses as loose evidence that might not represent the workforce. The figure concerns employees’ self-reports, not AECI benchmark growth.
METR said AI was not yet making highly significant, fully autonomous contributions to the most important stages of model training. Its estimate of AI’s contribution to faster progress therefore describes assistance, not a fully autonomous model-development process.
What METR examined
METR’s assessment ran from August 12 to October 29, 2026. Over the 11-week engagement, it reviewed company-shared benchmark data and other internal information and interviewed staff. It did not query unreleased models.
Anthropic could unilaterally redact nonpublic information from the public report. METR said those redactions obscured much of the evidence it found most informative, including parts of its analysis of AI’s contribution. The assessment thus draws on internal information, while the public account leaves substantial parts of that evidence unavailable for outside scrutiny.
