More capable models are better at in-context scheming