AI task capability improvement estimated to be twice as fast… Key evidence for Anthropic's evaluation not disclosed

Nonprofit AI evaluator METR reported that Anthropic’s models may be improving at task capability roughly twice as fast as before, based on an independent assessment covering August 12 to October 29 and benchmarked against data from March 2024 through October 2025. METR estimated the true acceleration could range from 1.3x to nearly 3x prior rates, and cautioned that it is unclear whether the same pace applies to broader tasks without clear success criteria. The finding matters because it offers a rare external read on how quickly Anthropic’s frontier models are advancing, even as the company withholds key internal evidence behind the evaluation.