With the Gemini 2.5 Flash retirement coming closer I’m scrambling to try and find a suitable replacement for 2.5 Flash for a realtime voice agent pipeline.
In Australia Gemini 2.5 Flash inference latency was ~300ms.
Gemini 3.5 Flash has inference latency of ~500ms, but additionally is no longer supposed in Australia, the nearest deployment being in Singapore which is +100ms per turn.
300ms > 600ms per LLM call may not sound like much but it is absolutely devastating in a latency sensitive system where the target for STT+LLM+TTS is ~700ms. Previously this was achievable with Gemini 2.5 Flash, but no longer with 3.5 Flash.
Additionally the ~600ms inference time has dramatically reduced Gemini’s usefulness for semantic end-pointing and turn precompute.
Losing Gemini 2.5 Flash is a major loss to APEC. It had the fastest inference speed of any model in the region.