The same thing they’re doing with Gemini, creating custom versions, is likely what they’ve done with Claude and OpenAI models as well. They’re likely evaluating all of them internally with employees all the time.