Google's Gemini 3.8 Live lets a voice agent reason and call tools without going silent

Gemini 3.8 Live and its Extended Thinking variant are native speech-to-speech models that run tool calls and reasoning in the background while audio keeps streaming. Priced per minute of audio rather than per token, they target the dead air in today's voice agents.

By Stackmaven

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, two native speech-to-speech models aimed at developers who want a voice agent that can keep working while it keeps talking. The voice quality is good, but the more interesting change is where the work happens. Both models can run tool calls and multi-step reasoning in the background while audio keeps streaming, which targets the artifact that makes most voice agents feel robotic: the silence while the system thinks.

What native speech-to-speech changes

Most production voice agents today are a relay. Speech-to-text feeds a language model, the model’s text answer feeds a text-to-speech engine, and every hop adds latency. The worst symptom of that pipeline is dead air: when the model needs to call an API or reason through a multi-step request, the conversation stops until the whole chain finishes. Gemini 3.8 Live is a single native speech-to-speech model, so audio in and audio out are one model rather than three stitched services. The part that changes how an agent behaves is asynchronous function calling: the model can say “let me pull that up” and hold the floor while a tool runs in the background, then fold the result into its reply. The Extended Thinking variant applies the same idea to reasoning, running configurable multi-step thinking in the background and narrating progress instead of going quiet. For a developer, the practical shift is that the awkward pause you currently design around, with filler audio or a spinner, becomes something the model can carry inside the turn.

Per-minute pricing changes the cost math

Gemini 3.8 Live is priced by audio duration, not tokens: $0.005 per minute of audio input and $0.018 per minute of audio output, per Google’s developer documentation. That is a real departure from the per-token pricing developers budget most LLM features against, and it maps cleanly onto the workloads voice agents actually serve, like support calls and phone assistants, where minutes are the natural unit. It also front-loads a planning question. Cost now scales with how long people talk, not how much text moves, so a rambling ten-minute call is expensive in a way a token-metered chat turn is not. Teams modeling a voice deployment should price against expected talk time and call length rather than prompt size.

What it means for developers

The launch reads as a build-versus-buy signal for anyone assembling a voice stack. Gemini 3.8 Live is available now through the Gemini API and Google AI Studio, and Google is shipping it through the real-time voice SDKs many teams already reach for, including LiveKit, Pipecat, Vercel, LangChain, Agora, Fishjam, and Vision Agents, which handle the media-streaming plumbing so the model is closer to a drop-in than an infrastructure project. Google also released Gemini 3.5 Transcribe alongside it, a speech-to-text model covering 85 or more languages with a reported word error rate of 4.0 percent streaming and 2.6 percent non-streaming, plus speaker labeling and custom vocabulary biasing. If you have already built the stitched speech-to-text and text-to-speech pipeline, the question this raises is whether the native model removes enough glue code and latency to justify a migration. If you are starting now, the calculus tilts toward the native path.

What to watch

The capability claims are Google’s own for now. It reports a top score of 82.6 on the Artificial Analysis Speech-to-Speech Quality Index and 68.6 percent on the agentic τ-Voice benchmark, but those are vendor-cited figures, and the meaningful test is independent evaluation on real call flows rather than a leaderboard position. The Extended Thinking model’s narrate-while-you-think behavior is also a bet on user experience: progress narration can feel natural, or it can feel like a stall with commentary, and which one it is will depend on how carefully teams tune it. The signals worth tracking over the next quarter are whether independent benchmarks reproduce the speech-quality and agentic numbers, whether per-minute pricing holds up against token-metered alternatives once real call volumes land, and whether asynchronous tool calling measurably cuts the latency workarounds developers currently ship. Stackmaven’s follow-up will revisit around December 15.

Sources cited