“Anthropic's stated position is that reasoning should be an integrated capability within a single model, not a feature of a separate model.”