Build1 publisher2 min readPublished
Google introduces two Gemini 3.8 Live audio models, one for high-volume chats and one for complex reasoning
Google positions one live model for high-volume conversation and the other for reasoning that continues while the user is still talking. The product page describes both without performance measurements, token costs or latency figures.
The Engineer · Build desk
What happened
- Google DeepMind introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two audio-focused models built for near real-time spoken conversation.
- The pair extends the Gemini Audio family past transcription, translation and text-to-speech, with one model aimed at high-volume voice traffic and the other at harder reasoning during a live exchange.
- Google says the Extended Thinking variant handles complex reasoning and can coordinate multiple agents on background tasks while keeping the conversation going.
- Google published no performance, cost or latency numbers for either model in the official material, so there is nothing there to compare them on operationally.
- Access is listed through Google AI Studio, the Gemini Live API, the Gemini API, the Gemini app and related Google services.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision A team has to name a default model per application before any comparison exists, so the choice rests on Google's positioning language, and the latency and quality figures have to come out of the pilot.
- cost Per dev.to, the page lacks the detail needed to say which model is cheaper for a given workload, so whoever owns the voice product carries the estimating work and the budget variance.
- constraint The two models are positioned for different jobs, so changing the default later touches the interaction design as well as the model string in the config.
- exposure SynthID is offered as the way to flag AI-generated or edited speech. Any organisation disclosing synthetic voice to customers builds the verification workflow and the policy around it.
Routing by complexity needs three numbers per model: the cost of a minute of audio, the added latency when the harder model reasons, and the quality delta on your own turns. The third was never going to come from a vendor page. Google describes Gemini 3.8 Live as optimized for high-volume, cost-effective use, and the page does not list a price for it [3][5]. dev.to, which wrote up the release, reports that the comparison table it reproduces shows Google's stated positioning [7].
Extended Thinking's orchestration claim is the one that changes what routing means [4]. If the model doing the hard work can keep talking, the choice is per application or per session, not per turn. In a real product that claim rests on the session mechanics. The session has to keep streaming audio while tool calls run. Barge-in cannot kill the background task when it cuts the speech. And when a background task fails, the failure needs to surface in the dialogue instead of leaving the model to improvise.
AI Studio is the exploration surface; the Live API and the Gemini API are where an integration lands [12]. Run the same recorded script through both models on the API path and you get the latency delta as a measurement: time to first audio, then time to a completed answer. The latency delta decides whether Extended Thinking is usable for a live caller. Cost per conversation stays an estimate.
The live models sit beside the task-specific ones. Google's page also lists Gemini 3.5 Transcribe, Gemini 3.5 Live Translate and Gemini 3.1 Flash TTS [9], which with the two new entries makes five audio models across three version numbers [11]. A transcribe-translate-speak pipeline assembled from those parts pins 3.8, 3.5 and 3.1 at the same time.
So the sequence that survives the gaps is an ordinary one: default to the cheaper conversational model, log the turns where it falls short, and escalate to Extended Thinking on measured need. That pilot sets the escalation threshold and the cost model.
What to watch
- A published per-minute or per-token price for either 3.8 Live model. That price would turn pilot latency numbers into a unit cost.