Also Claude/Fable models are quite bad at instructions following: https://artificialanalysis.ai/evaluations/ifbench