Here's how the 1M token output (long decode) works: https://ai.google.dev/gemini-api/docs/long-decode