Gemini 3.8 Live puts voice agents into production

Google's voice stack now has two production tiers. Gemini 3.8 Live is the cost-efficient one meant to run at volume; Gemini 3.8 Live Extended Thinking reasons across several steps and keeps talking while it does. Both are native speech-to-speech, so audio never round-trips through a separate transcription and text model. They went live on September 15, 2026.

What is actually new

The headline capability is asynchronous function calling: the model fires a tool or API request in the background and keeps the conversation moving, narrating progress instead of going silent. It handles 97 languages and can switch between them mid-conversation, per Google's announcement post. Google also claims near real-time visual context and tighter accuracy on alphanumeric strings, the confirmation-code problem that trips up most phone agents.

Benchmarks, and the gap inside them

Extended Thinking takes first place on the Artificial Analysis Speech to Speech Quality Index with 82.6. Google reports 68.6% on τ-Voice for agentic task completion, 35.1% on Sierra's τ-Voice-banking, and 97.7% on Big Bench Audio. The smaller 3.8 Live lands second on Speech Agent Arena, a human-preference ranking.

Read those together and the picture is more sober than the top ranking suggests. Audio reasoning is close to saturated at 97.7%, while the banking workflow, the one that resembles a real support call wired to real back-end systems, sits at 35.1%.

What it costs and where to get it

MarkTechPost converts the $3 per million input tokens and $12 per million output tokens into $0.005 per minute of audio input and $0.018 per minute of audio output. Access runs through the Gemini Live API and Google AI Studio, and there is no self-hosted or open-weights option. Enterprise customers get a private preview in Gemini Enterprise. Consumers meet 3.8 Live in Search Live, and Extended Thinking in Gemini Live, Gmail and Keep, plus Google Workspace on Pro and Ultra plans.

Launch integrations cover Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents. Every generated audio clip carries a SynthID watermark.

A same-day look from outside Google

Simon Willison shipped a dependency-free browser tool on release day. It opens a WebSocket to Google's generative language endpoint and uses the Web Audio API for both capture and playback. You pick a model and a voice preset, optionally add a system prompt, and can cut the model off mid-sentence. He reads the release as a similar shape to OpenAI's GPT-Live family, and cautions that transcripts may include speech that was interrupted before it ever played.

What is still unreported

None of the sources checked here give a latency figure in milliseconds. The quoted “impressive latency” is partner testimony, not a measurement. The context window for either model is likewise absent from these sources. Treat both as unpublished rather than unimpressive.