Gemini 3.8 Live Adds Reasoning to Real-Time Voice AI

Voice interfaces have spent years stuck in an awkward trade-off. Make the model fast enough to hold a natural conversation and it cannot think very hard; give it room to think and the pauses break the illusion of talking to something. On September 15, 2026, Google announced two models that attack that trade-off directly: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, both native speech-to-speech models aimed at production voice agents.

• Two models launched September 15, 2026: Gemini 3.8 Live for scale and cost efficiency, and Gemini 3.8 Live Extended Thinking for multi-step reasoning
• Extended Thinking ranks first on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6, and hits 97.7% on Big Bench Audio reasoning
• Both models detect and switch between 97 supported languages mid-conversation, without being told to
• Available now to developers through the Gemini API and Google AI Studio, with rollout across Gemini Live, Search Live, Gmail, Docs and Keep

What Makes Gemini 3.8 Live Different From a Voice Wrapper?

Most voice assistants are a pipeline: speech-to-text, then a text model, then text-to-speech. Every hop adds latency and throws away information — tone, hesitation, the sound of someone trailing off. A native speech-to-speech model skips the intermediate transcript entirely and works on audio end to end, which is why these systems can interrupt gracefully and match a speaker's pacing.

Gemini 3.8 Live is the volume model in the pair. Google positions it for scale and cost efficiency, combining conversational fluency with visual grounding — it can take camera or screen input and reason about what it sees while the conversation continues. It placed second in Speech Agent Arena, according to Google's announcement.

Extended Thinking is the more interesting engineering story. Rather than going silent while it works through a hard problem, the model reasons and speaks at the same time, filling the gap with natural verbal acknowledgments like "Let me check that…" while a tool call runs underneath. Anyone who has sat through four seconds of dead air waiting on an agent will recognise why that matters. Google reports 68.6% task completion on the τ-Voice agentic benchmark and 35.1% on Sierra's banking benchmark for the Extended Thinking variant.

How Good Are the Benchmark Numbers?

The headline figure is the 82.6 score on Artificial Analysis' Speech to Speech Quality Index, which Google says puts Extended Thinking at the top of that leaderboard. Paired with 97.7% on Big Bench Audio — a reasoning test delivered as audio rather than text — the picture is of a model that is not simply transcribing well but holding the thread of an argument delivered out loud.

These are the vendor's reported figures against public benchmarks, and the usual caution applies: leaderboard position is a snapshot, and voice quality is notoriously hard to capture in a single number. What is independently checkable is the shape of the release — two tiers, one tuned for cost and one for reasoning depth, which is the same pattern the industry has settled into for text models.

The 97-Language Switch Is the Underrated Feature

The detail most likely to show up in real deployments is automatic language transition. The models detect and move between 97 supported languages mid-conversation, without a mode toggle or a restart. For a support line in a multilingual city, or a family where three generations speak two languages in the same sentence, that removes a configuration step that most systems still demand up front.

Background tool execution rounds it out. The model can fire API calls and keep talking, rather than freezing the dialogue while a lookup resolves — the difference between a voice agent that feels like software and one that feels like a person checking something.

Where You Can Use It Today

Developers get both models through the Gemini API and Google AI Studio immediately. Enterprise customers reach them through a private preview in Gemini Enterprise. On the consumer side, Google is routing the models into Gemini Live, the Search Live experience in AI Mode, and Workspace surfaces including Gmail Live, Google Docs Live and Google Keep Live, per 9to5Google's breakdown of the rollout.

For readers tracking how fast the real-time voice tier is moving, this lands days after OpenAI's own voice-agent pricing shift — see our coverage of the GPT Live 1 API for voice agents — and follows Google putting Gemini on the Windows desktop earlier this month. More model releases are collected on our artificial intelligence coverage page.

The broader read: voice is no longer the shallow end of the model pool. When a speech-native model can hold a reasoning chain, call tools mid-sentence and switch languages on the fly, the interface stops being a demo and starts being infrastructure.

Sources: Google — Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking — September 15, 2026; 9to5Google — September 15, 2026.