Google DeepMind has released two new audio models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The company describes them as its most advanced live dialogue models yet.
Both are native speech-to-speech. There is no intermediate step of transcribing and re-synthesising; audio goes in and audio comes out.
The gap being targeted
The release aims at a specific shortfall: voice agents that can reason and execute tools without breaking the flow of conversation.
In setups so far, an agent would fall silent while calling a tool and the conversation would break. That is why the Extended Thinking version ships as a separate model: to carry out work needing longer deliberation without cutting off the exchange.
The Extended Thinking split also creates a choice on the usage side. In a simple appointment bot, longer deliberation is wasted cost; in a flow that checks an order and starts a return, it is necessary. Wiring the two models to different flows in the same product is possible.
The price difference
The striking line is the price. Google charges 1.38 dollars per hour of voice conversation, markedly below OpenAI's GPT-Live-1.
| Criterion | Gemini 3.8 Live | GPT-Live-1 |
|---|---|---|
| Hour of conversation | 1.38 dollars | markedly higher |
| Leaderboard | Tops the speech-to-speech list | — |
| Naturalness | — | More natural thanks to full duplex |
The last row matters: OpenAI's model is expected to still sound more natural because it runs full duplex. The choice is not one-dimensional; there is a trade-off between cost and conversational fluidity.
Topping the leaderboard is not sufficient on its own either. Speech-to-speech rankings mostly measure English; how the same model behaves in another language remains a separate question.
Where it is available
Both models are live through the Gemini Live API and Google AI Studio. They are offered as hosted models, and since no open weights were published there is no option to run them on your own servers.
That settles the first question to ask when building a call centre setup that handles personal data: where is the audio processed and how long is it kept. The price advantage changes meaning depending on that answer.
The rest of the family
The new models sit on top of the audio family Google expanded last month with Gemini 3.5 Transcribe. The company is building a complementary line rather than isolated pieces: transcription, live dialogue and extended deliberation are positioned as separate products.
For a non-English deployment, what decides the outcome is language before price. However low the hourly cost falls, if the misunderstanding rate stays high in your language a voice agent does not work in customer service; that is why measuring on your own recordings has to come before the decision.