I am developing a commercial AI Voice Agent based on Vertex AI, which critically depends on the Gemini 2.5 Live API. This version offers architectural features essential to my service: native grounding, proactive audio, and asynchronous function calling.
I noticed that version 2.5 Live will reach its end-of-life on December 13, 2026. To prepare, I began migration testing on subsequent versions (3.1 and 3.5 Live), only to find these features missing. The removal of native grounding represents a critical roadblock for me: the accuracy of my Voice Agent’s responses relies entirely on this integration. The lack of feature parity between the deprecating version and the newer ones jeopardizes the continuity and quality of the service I offer my clients.
Troubleshooting steps attempted
-
Prompt Emulation: I attempted to replace native grounding by injecting context through advanced prompt engineering in versions 3.1 and 3.5 Live. The result is insufficient for a production environment, as it significantly increases the risk of hallucinations compared to the native feature.
-
Documentation Analysis: I reviewed the release notes and Vertex AI documentation to identify alternative APIs, ongoing beta tests, or official workarounds to restore proactive audio and asynchronous function calling, but found no viable solutions.