Gemini 3.8 Live ships native speech-to-speech voice agents
Google's two new Live models reason and call tools mid-conversation at $3/$12 per 1M tokens; a 2.6% WER transcriber and an open Suno rival ship alongside.
Copy markdown
Reasoning while it talks
The Live models fire tool and API calls in the background while they keep streaming audio, so an agent can look something up mid-sentence without dead air — plus real-time visual grounding to react to what it sees.
The price that makes it production-ready
Audio runs $0.005/min in and $0.018/min out — roughly $3 and $12 per 1M tokens. Cheap enough to put a live voice agent in front of real customers, not just a demo.
It tops the speech-to-speech board
Gemini 3.8 Live sits #1 on Artificial Analysis' Speech-to-Speech Quality Index at 82.6, with 68.6% on the τ-Voice agentic task benchmark and 97.7% on Big Bench Audio. It handles 97 languages and switches between them mid-call.
Two tiers, and it drops into your stack
Plain Live is tuned for low latency and cost; the Extended Thinking variant reasons in the background while speaking. Both are live on the Gemini API and AI Studio with ready-made support for LiveKit, Pipecat, LangChain and Vercel.
Gemini 3.5 Transcribe rides along
A dedicated speech-to-text model at 2.6% word error rate (4.0% streaming) across 85+ languages, with custom-vocabulary biasing and sub-second captioning on files up to an hour. Note: all three are hosted-only, no open weights.
Elsewhere: an open Suno rival you can run tonight
OpenMOSS dropped YuE2-3B, a 3B open-weights music model (~7.3GB) that generates full songs from lyrics plus a style prompt and beats Suno v5 on its makers’ own benchmark. Weights are on Hugging Face for local runs.