Okay, the new native model `gemini-2.5-flash-native-audio-preview-12-2025` is night and day better at instruction following than the previous native model. I have a usecase where it has to be a strict examiner examining code. And it actually does it quite well. Always ensuring that the syntax is correct.
The downside is that it feels more robotic to talk with, where the previous model was quite upbeat, bubbly and talkative. This new model answers more in short sentences and only exactly what it thinks is relevant to say.
But again Big improvement, super happy about this