AI agents reportedly vary widely in how they learn from experience

A post describing a Meta paper says researchers kept model weights fixed while agents revised tools and notes after rounds of chess, Go, Hex and NetHack. It reports that Claude Fable 5 rose from 37.5% to 73.3% on held-out hard-chess games. Across chess, Go and Hex, GPT-5.6 Sol gained 298 points per $1,000 of learning cost, versus 57 for Claude Fable 5. The post argues that evaluations should track learning gains and their cost.