An index measures how far AI is from improving itself

On 15 October, Vals AI updated its index of recursive self-improvement, the idea of an AI developing its own successors without human help. It has models carry out four AI research tasks on their own, with limited time and computing power. A score of 0.5 means matching a reference result, whether published, human or from another model.

The best figure it publishes for the overall index is Claude Fable 5.1's 0.36, and no model reaches 0.5, although Claude Opus 5.5 beat the reference on one of the tasks: training a small language model. Vals warns that these are single runs on narrow tasks and that the references don't measure the best that humans can do.