Google's Gemini 3.8 Live Pushes Voice Agents Toward Production
Gemini 3.8 Live is now available in Google's Gemini API and AI Studio, and it ships alongside a second variant, Gemini 3.8 Live Extended Thinking, that carries multi-step reasoning through an open voice session. Google published the launch on September 15, 2026, pricing audio at $0.005 per minute for input and $0.018 per minute for output. The company states that the model ranks first on the Artificial Analysis Speech to Speech Quality Index with a score of 82.6.
A speech-to-text model accompanies the release. Gemini 3.5 Transcribe handles more than 85 languages. Google reports that it misreads 4.0% of words when transcribing live audio and 2.6% of words when it works from recorded audio, and it accepts custom vocabulary biasing of up to 1,000 terms. Google applies SynthID watermarking to every audio stream the Live models produce.
What Gemini 3.8 Live Adds to the Stack
The core capability is conversational continuity with visual grounding. An agent built on Gemini 3.8 Live can reference what a camera or screen is showing while the user talks, which points toward field service, accessibility tools and retail assistants rather than plain chat.
Multilingual support spans more than 97 languages, with automatic detection and switching mid-session. For a support line serving mixed-language callers, that removes the routing step that would otherwise send each caller to a language-specific bot.
The language counts do not line up across the two models. Gemini 3.8 Live covers more than 97 languages while Gemini 3.5 Transcribe covers more than 85, so a pipeline that pairs live dialogue with transcription loses coverage wherever the transcript model stops. Teams building multilingual agents should size their plans against the narrower figure.
Asynchronous function calling is the less visible change. Background tasks such as a booking lookup or a database write run without interrupting speech, so the user does not sit through silence while a tool call resolves. Gemini 3.8 Live Extended Thinking applies the same principle to harder requests by keeping multi-step reasoning inside the conversation.
The two features address different problems. Extended Thinking buys depth: an agent weighing several constraints can reason before answering instead of handing the request to a separate reasoning model and a visible pause. Asynchronous function calling buys flow: work that need not block the reply stops blocking it.
Splitting fast dialogue from deep reasoning into two model names is itself a product choice. A single model that answers instantly and reasons carefully would have to absorb the latency somewhere. Two models let developers choose per call, sending routine turns to Gemini 3.8 Live and difficult ones to Extended Thinking, at the cost of routing logic to write and a second set of behaviours to test.
The 3.8 designation also puts the Live models on the same generation track as Gemini 3.8 Flash, which Google lists as available in the Gemini API. Sharing a version number simplifies the choice for teams already building on the 3.8 family.
| Model | Role | Availability |
|---|---|---|
| Gemini 3.8 Live | Near real-time voice dialogue with visual grounding | Gemini API and AI Studio; private preview in Gemini Enterprise; Search Live and Gemini Live for subscribers |
| Gemini 3.8 Live Extended Thinking | Multi-step reasoning inside an open voice session | Same channels as Gemini 3.8 Live |
| Gemini 3.5 Transcribe | Speech-to-text across 85+ languages; 4.0% word error rate when streaming | ai.studio/live |
The Unit Economics of a Voice Minute
Output audio costs 3.6 times input audio at list price, so the balance of talking matters more than the headline rate. A ten-minute session split evenly between user and model comes to roughly $0.115, and one million such sessions a month would run about $115,000 in audio cost before text tokens, tool calls or infrastructure. An agent that speaks more than it listens pays a premium for the habit.
| Session length | Estimated audio cost (even talk split, list price) |
|---|---|
| 5 minutes | $0.058 |
| 10 minutes | $0.115 |
| 30 minutes | $0.345 |
| 60 minutes | $0.69 |
Per-minute billing changes how product teams design as well. A voice feature that keeps a session open while the user thinks costs money in a way a text box does not, so timeouts, turn detection and session length become budget decisions rather than pure interface decisions.
Gemini 3.5 Transcribe uses a different meter: $3 per million input tokens and $12 per million output tokens. Token billing gives developers finer control over what they submit, and it makes forecasting harder, because the token count depends on audio length, language mix and how much vocabulary biasing is applied.
The distance between the two accuracy figures shows what live output costs. Streaming adds 1.4 percentage points of word error over batch processing, the accuracy a model gives up when it must commit to words before a sentence ends. Live captioning usually justifies that trade. Compliance transcripts do not.
Transcription accuracy carries a downstream cost the percentage hides. A 4.0% word error rate on a twenty-word instruction leaves roughly one word wrong per request, and a voice agent that acts on a misheard address, quantity or date produces a failure the user has to notice and correct. Reliability targets for agents that execute tasks have to exceed the transcription rate.
Google named LiveKit, LangChain, Agora and Vercel as integration partners for media streaming infrastructure. That list describes the intended build path: a streaming layer, an orchestration framework and the model, with less transport code for the developer to own. Smart transcription mode in Gemini 3.5 Transcribe, which strips filler words and self-corrections, supports the same workflow by handing cleaner text to downstream agents.
Where the Trade-offs Sit
Enterprise access is the first constraint. Gemini Enterprise gets a private preview rather than general availability, so regulated buyers cannot yet ship production agents on the model without engaging Google directly. Consumer surfaces, where Search Live and Gemini Live reach subscribers, will move faster.
Latency for the Extended Thinking path is not published. The 82.6 score measures conversational quality, not the time a reasoning turn adds to an exchange, and that number decides whether an agent feels responsive or stalls.
Billing for reasoning is also unstated. Google has not said whether tokens consumed by Extended Thinking are charged at the same audio rates or metered separately, which leaves cost models with an unknown variable at the point where usage is hardest to predict.
The same model reaches users through two different cost structures. Developers pay per minute of audio, which scales with usage and lands on the application's budget. Subscribers to Search Live and Gemini Live pay a flat fee, so Google absorbs the inference cost and prices the bundle on engagement rather than minutes. The retail price of voice therefore tells a competitor little about the marginal cost of running it.
Watermarking is a detection aid rather than a complete control. SynthID is applied to every generated audio stream, which helps platforms screen synthetic speech, though detection depends on the watermark surviving compression and re-encoding. Teams should treat it as one signal alongside other verification.
The practical verdict: API and AI Studio builders can evaluate Gemini 3.8 Live now, and per-minute pricing makes cost modelling tractable at moderate volume. Enterprise buyers should read the private preview as a prompt to open procurement conversations rather than a deployment date. The decisive test is the Extended Thinking turn, because seconds added to a reasoning reply will shape user retention more than a quality index score.
Why this matters
Voice is shifting from demonstration to line item. Per-minute pricing, non-blocking tool execution and watermarking cover three requirements a production voice product needs: predictable cost, backend work that does not stall the conversation, and an auditable audio trail. The competitive question has moved from whether a model can hold a conversation to whether it can do so at a cost and latency that hold up under real traffic.
Sources
Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe
Release notes | Gemini API | Google AI for Developers
Models | Gemini API | Google AI for Developers
Live API capabilities guide | Gemini API | Google AI for Developers
Gemini 3.8 Flash — Google DeepMind
Turn your voice into action with new productivity features in Gemini Live
Gemini 3.1 Flash Live: Making audio AI more natural and reliable
Gemini Audio — Google DeepMind
Related Articles
- Google's Gemini Live API Pricing Turns Voice Agents Into a Metered Commodity [Update]
- Google Launches Gemini 3.5 Live Translate for Continuous Real-Time Speech Translation
- Interactions API Now Default Gateway for Gemini Agents
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.