Comment by chrisjj

Sure. I was separating the model - which Anthropic promises not to downgrade - and the "thinking time" - which Anthropic doesn't promise not to downgrade. It seems the latter is very likely the culprit in this case.