Meta Superintelligence Labs proposes 'agent plasticity' metric for AI self-improvement efficiency

Researchers from UC Berkeley, Meta Superintelligence Labs and other institutions propose agent plasticity, a metric defined as gain on held-out tasks per dollar spent on learning, measured with model weights frozen and each run starting from a fresh context.

The paper reports that on chess, Go and Hex, Claude Fable 5 reaches the highest final score while GPT-5.6 Sol gains the most per dollar. On NetHack, the paper reports that only Claude Opus 5.5 improves significantly.

The finding is reported from a single paper, described in a social media post by Elvis Saravia, and the summary does not detail the benchmark sizes or cost calculations behind these rankings.