I don't know what they did but opus is really good while Astra/Sol are comparably bad. For my own tests and taste this might be the first time (fable perhaps excluded) where claude models are better than openai, since codex 5.3.