The latest version of a high-end AI model for agentic reasoning outscored Fable's newest release on a recent industry benchmark. Claude Opus 5.5 (xhigh) recorded a 31.2% score on the WeirdML v3 test, surpassing the 26.0% result of Fable 5.1 but trailing the 42.2% performance of GPT 6 Astra.

The xhigh model is priced at less than half the cost of the previous Opus 5 iteration. The WeirdML v3 benchmark consists of 11 complex hand-made tasks that require models to develop machine learning and data analysis pipelines to handle unfamiliar data with limited feedback or unspecified goals.