With gemini-1.5-flash-001 being retired, there’s not a single Stable(GA) Gemini model, either 1.5 or 2.0 that can accurately refer to audio sections by timestamps (easily verifiable in AI studio or via API using the default prompt in the example with any audio), an issue that has been brought up in this forum several times.
If the trend of this working while models are in Preview/Experimental then severely downgrading performance after GA continues, it would be nice to add a disclaimer for the sake of transparency that this feature is intended for Preview/Experimental models only.