METR finds that Anthropic improves AI 1.9 times as fast as before

METR, an external group, did a test at Anthropic with high access, from 12 August to 29 October 2026. METR tells that Anthropic improves its AI models on well-specified tasks at about 2 times the rate before October 2025. METR thinks that AI use in the research of Anthropic causes most of this. METR is not sure of the estimate: the rate can be as low as 1.3x or as high as 3x.

How it works

METR got internal data from Anthropic: training metrics, use of computer power, benchmark scores and staff interviews. It tracked AECI, the version of Anthropic of the Epoch Capabilities Index. Staff tell that AI helps them. At this time, AI does not do the research without a person.

The numbers

After October 2025, the AECI rate was 1.9 times the rate before, with a 95% confidence interval of 1.5 to 2.8 times. That is about two years of development in one year. METR is sure that more computer power for training was not the cause.

What is not known

Anthropic removed most of the important data. It tells that it removed data that can make development fast for other AI developers. METR and Anthropic researchers are not sure how to read the other data. METR saw no sign of a limit on development in the next 12 months, but its view changed some times.

Insight: A lab that increases its rate shows less of its data

In October 2026 Anthropic stopped live internet access for its internal AI tests until it can monitor and control its agents. METR tells that Anthropic moves about 2 times as fast, and removed most of the important data. This reads as a pattern: when a lab increases its rate, it removes more data, and external groups become the only check. The next move to know is if other labs give METR the same access.