Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking: Google DeepMind Splits Voice AI Into Two Speeds

Google DeepMind has introduced two new members of its Gemini Live family, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, splitting conversational AI into a fast, scalable tier and a slower, deeper-reasoning tier. The distinction matters because voice interfaces have historically forced a choice between responsiveness and accuracy, and these two models are positioned as separate answers to that same question rather than a single compromise.

Two models, one underlying tradeoff made explicit

Gemini 3.8 Live is described by Google as built for scale and cost efficiency, combining conversational intelligence, fluid dialogue, and visual grounding, with visual inputs processed in near real time. Gemini 3.8 Live Extended Thinking is built for high-complexity tasks, supporting multi-step reasoning and the ability to reason and speak simultaneously rather than pausing to think before responding. That last detail is worth sitting with: instead of the usual latency penalty that comes with deeper reasoning, Extended Thinking is designed to keep talking while it works through a harder problem. Whether that holds up under sustained use is a separate question from what Google claims, but the architectural intent is clear enough to shape how a developer should think about picking between the two models.

What the benchmark numbers actually say

Gemini 3.8 Live Extended Thinking placed first overall on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6, while the standard Gemini 3.8 Live ranked second in the separate Speech Agent Arena. On Sierra's τ-Voice benchmark, Extended Thinking scored 68.6%, and on the narrower τ-Voice-banking benchmark it scored 35.1%, a gap that on its face suggests the model performs considerably better on the general benchmark than on the banking-specific one, though Google has not published a breakdown explaining why. On Big Bench Audio, Extended Thinking scored 97.7%. These numbers describe relative standing on specific third-party tests, not a general claim of readiness for any particular deployment, and a developer evaluating either model for a real workload should treat the benchmark scores as a starting point for their own testing against their own task, not a substitute for it.

Where the audio comes from and how you'd know

Google states that audio generated by its AI products, which would include output from both new models, is watermarked with SynthID. That's a provenance mechanism, not a content filter, and it matters for anyone building a product where a user might reasonably ask whether they're hearing a synthetic voice. Separately, Gemini 3.8 Live can execute tools and API calls in the background while continuing a conversation, meaning the model doesn't have to go silent while it waits on an external system to respond. Combined with support for automatic transitions between 97 languages mid-conversation, the practical shape of these models is a voice agent meant to keep a conversation moving even when it's doing other work underneath, rather than one built purely to answer isolated single-turn queries.

The developer platforms Google is naming, and what that does and doesn't mean

Google has listed Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents as developer platforms for the Gemini Live API. Google's own materials describe this as a list of platforms, not a claim that each one ships turnkey integration or that a developer already using one of them gets Gemini Live support automatically; anyone building on one of these platforms would need to check that platform's own documentation for how, or whether, Gemini 3.8 Live and Extended Thinking are exposed there today. The actual point of entry Google specifies is simpler and more direct: both models are available through the Gemini API and through Google AI Studio, which is where a developer would go to get credentials, test a prompt or a live session, and start building without going through any of the seven named platforms at all.

Deciding which tier fits a real project

For anyone weighing whether to build against Gemini 3.8 Live or Extended Thinking, the practical distinction to test for is not which model scores higher on a leaderboard but which one your task actually needs. A high-volume, cost-sensitive application, such as a customer-facing voice interface fielding large numbers of short exchanges, is the scenario Google frames Gemini 3.8 Live around. A task that requires the model to work through several dependent steps before it can give a useful answer, the kind of situation the τ-Voice and Big Bench Audio benchmarks are meant to approximate, is where Extended Thinking's stated design applies. Testing both against your own use case through the Gemini API in Google AI Studio, rather than assuming the benchmark rankings transfer directly, is the only way to know which tradeoff actually fits what you're building. That kind of hands-on evaluation, comparing model behavior against a concrete workflow instead of a spec sheet, is exactly the exercise NextLUCA works through with teams building applied AI systems, whether the starting point is a single agent prototype or a full production integration.

Specs

• Supported languages: 97, with automatic transitions during conversation
• Speech to Speech Quality Index (Extended Thinking): 82.6, ranked first overall
• τ-Voice (Extended Thinking): 68.6%
• τ-Voice-banking (Extended Thinking): 35.1%
• Big Bench Audio (Extended Thinking): 97.7%
• Access: Gemini API and Google AI Studio, for developers

Sources

• Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — Google DeepMind