METR estimates faster progress in Anthropic AI models

On November 5, 2026, Model Evaluation and Threat Research (METR) published an assessment estimating that Anthropic’s model capabilities on well-specified tasks were advancing about twice as fast, as of October 2026, as during March 2024–October 2025. METR said the true pace could plausibly be closer to 1.3x or 3x. Its separate AECI analysis estimated a 1.9x acceleration in the fitted benchmark trend after October 2025. A third figure, about 12x, came from informal employee polls about internal Model 4. METR’s assessment treats these as different measurements, not one general measure of AI progress.

For context on the policy debate, NeoTeo previously covered Dario Amodei’s proposal to slow frontier AI development without stopping it.

What the 2x and 1.9x estimates measure

METR’s roughly 2x figure is its broader judgment about the rate at which Anthropic models were gaining capability on well-specified tasks, compared with progress from March 2024 through October 2025. METR said a rate closer to 1.3x or 3x would also be plausible.

The 1.9x figure comes from a fitted trend in AECI, Anthropic’s version of the Epoch Capabilities Index. AECI combines scores from internal and external benchmarks. METR estimated that the index rose by 15 points per year before October 2025 and 29 points per year afterward. A 15-point increase corresponds to a doubling in benchmark difficulty.

The numbers make sense only alongside what each one measures: task capability, a benchmark trend, or employees’ reported uplift.

What the 12x figure refers to

The 12x figure came from informal Slack polls asking Anthropic employees about the AI uplift they experienced while using internal Model 4 in October 2026. It is a self-reported average, not the 2x judgment about task capability or the 1.9x fitted AECI trend. METR described the poll evidence as very loose and said it might not represent employees more broadly.

How much of the acceleration does METR attribute to AI?

METR’s subjective estimate was that perhaps 80% of the acceleration as of October 2026 came from AI use in research and development. It described that attribution as substantially uncertain and said non-AI factors likely also contributed to progress in late 2025. METR said it had convincing evidence that the acceleration was not caused by a faster rate of growth in training compute, but did not identify one alternative cause with comparable confidence.

METR described AI as primarily assisting research and development. It said AI had not yet made highly significant, fully autonomous contributions to the most important parts of model training, though researchers reported early signs that more autonomous contributions might occur soon.

How METR built the trend estimate, and its limits

For its main AECI analysis, METR fitted piecewise linear trends with one estimated breakpoint, placed on October 7, 2025. Its robustness checks produced estimates from 1.1x to 2.9x under alternative link-function assumptions, and from 1.9x to 2.2x under tested choices of model snapshots and benchmarks. The 95% confidence interval accounts for uncertainty in AECI scores, not uncertainty about which model snapshots entered the fit.

METR also found no convincing evidence of an algorithmic, training-compute, experiment-compute or data bottleneck that would take effect within 12 months. It said its view changed during the assessment and that this bottleneck judgment remained uncertain.

Anthropic had unilateral redaction rights over non-public material in the public report. METR said Anthropic redacted a majority of the evidence it considered most informative for its risk conclusions, limiting public scrutiny of the assessment’s full basis.