Anthropic report says Claude drives most of its AI research speedup

Anthropic let an outside group check whether its AI is now helping build its next AI. They found it is, and progress is running at about double the old speed. The detail buried at the end of their report is that most of the checking was done by the same AI. The group is METR, a nonprofit that tests AI models for risk. They spent 11 weeks inside Anthropic with access to internal data, unreleased models and staff interviews. Their conclusion is that Anthropic packed about two years of progress into the last twelve months. They ruled out the easy explanation, which would be buying a lot more chips, and their best guess is that roughly 80% of the speedup comes from Anthropic's researchers using Claude to do the work. Then you get to the section where METR asks whether its own audit can be…

METR says Anthropic has probably hit the 2x Responsible Scaling Policy acceleration threshold, and may now be running as high as 3x! 'We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.'

SITUATION DETECTED: METR estimates Anthropic is improving model capabilities at around 2× the March 2024–October 2025 rate, and may be running as high as 3×, with 80% of the current acceleration directly due to AI assistance, not compute scaling.

METR estimates Anthropic’s model benchmark scores are improving 1.9 times faster than before, and says AI may be driving about 80% of the current acceleration, though that guess is highly uncertain.