Google's Voice AI Now Thinks While It Talks
Gemini 3.8 Live and its Extended Thinking sibling handle complex reasoning, language switching, and background tasks mid-conversation.
By The ai-compute Desk
Google has shipped two new voice models — Gemini 3.8 Live and Gemini 3.Both are available now through the Gemini API, Google Workspace, and the Gemini app. [1]
The distinction between the two models is deliberate. Gemini 3.8 Live is built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding. Gemini 3.8 Live Extended Thinking is built for high-complexity tasks, with increased intelligence and multi-step reasoning. [1]
On the benchmark circuit, Extended Thinking captured the top spot on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6. It leads agentic task completion with 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark, and scores 97.7% on Big Bench Audio. [1]
The base model holds its own commercially: Gemini 3.8 Live secured second place in the Speech Agent Arena among users, and on ServiceNow's EVA-Bench — a benchmark for evaluating voice agents — the models push the Pareto Frontier for complex workflows by successfully balancing accuracy with conversational quality. [1]
The harder engineering problem was never raw accuracy — it was keeping the conversation alive while the model does real work in the background. That tension between thinking and talking has haunted voice AI since its inception.
Google's answer is a dual-track architecture. The model executes tools and API calls in the background while continuing the conversation, so it can acknowledge requests and keep chatting while tasks finish. Extended Thinking goes further: it reasons and speaks simultaneously, using early verbal cues like 'Let me check that…' to acknowledge prompts naturally, and live progress narration to walk users through multi-step background tasks as they progress. [1]
Visual context is threaded throughout. Gemini 3.8 Live processes visual inputs in near real-time, and it automatically detects and transitions between 97 supported languages mid-conversation — no manual toggle required. [1]
Developers get the infrastructure plumbing through platforms including Agora, Fishjam, LiveKit, Pipecat, Vercel, and Vision Agents, which manage real-time media streaming so builders can focus on the user experience. Enterprise partners Salesforce, Genspark, and Lumeris are already in early access, with Google citing the models' latency, fluidity, and tool-calling as the main draws. [1]
Voice interfaces have long been the poor cousin of text — capable enough for simple queries, brittle under anything resembling a real workflow. What's shifting now is the expectation layer: users want to hand off complexity, not just ask questions.
The base model powers step-by-step troubleshooting inside Search Live. All audio output is watermarked with SynthID — an imperceptible marker woven directly into the audio to keep AI-generated content detectable. [1]
Voice AI that can narrate its own reasoning while managing asynchronous tasks in the background is a meaningful jump in practical utility. Whether it translates to durable daily use — rather than impressive demos — will depend on how reliably these models hold up when the workflows get genuinely messy. [1]
For tasks that require deeper reasoning, 3.8 Live Extended Thinking reasons and speaks simultaneously.Gemini Audio Team