METR, an independent AI evaluation group that spent eleven weeks inside Anthropic this fall, estimates the company is now improving its models on well-specified tasks at about twice the rate it managed between March 2024 and October 2025. Its best guess is that around 80% of the acceleration is due to AI itself.

The measure is AECI, Anthropic's version of the Epoch Capabilities Index, which scores models across a set of benchmarks. METR puts the speedup at 1.9x, with a 95% confidence interval of 1.5x to 2.8x: about two years of historical progress in the last twelve months. The figure moves with the method, from 1.1x to 2.9x, and falls to 1.4x when models from OpenAI, Google and xAI are added. Anthropic's own published figure is 1.7x.

The supporting evidence is soft. Of eleven researchers whose interview answers could be coded, five said or leaned yes to a speedup of 2x or more, four were unclear and two said no. The larger worry is compounding. METR found no convincing bottleneck in algorithms, compute or data over the next 12 months, and concludes there is a significant risk that progress gets faster still.

That conclusion is hard to check. Anthropic held unilateral redaction rights and used them heavily: the bottleneck analysis, the algorithmic progress estimates, the split of progress between compute, data and algorithms, and the forecasts for future models are all missing from the public report. Anthropic says the cuts remove information that could speed up other developers. METR says much of what it saw had several plausible readings, and that its staff and Anthropic's researchers disagree on how to put the evidence together.

For investors, the open question is whether the hundreds of billions committed by Anthropic, OpenAI, Meta and Google buy real capability or just more compute. An outside estimate of a doubling gives defenders of the build-out something to point to, Anthropic's $895bn of spending commitments among them. But it rests on one lab's benchmarks, and Anthropic's newest model, Model 4, already scores above 50% on all 15 it was tested on, which makes the next jump harder to measure.

Our view: a lab whose models help build their successors, and which decides what outsiders may see of the result, should expect scepticism before credit.

© Raydn · raydn.us