Gemini 3.8 Live: voice agents that call tools mid-sentence

Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, and the interesting part is not the benchmark. It is that the model executes tools and API calls in the background without stopping the conversation. For anyone running a voice agent on a real phone line, that is the difference between a bot that sounds like a bot and one that sounds like a person checking a screen.

What actually happened

Two models, per Google's announcement: a cost-efficient Gemini 3.8 Live, and Extended Thinking with configurable reasoning effort for harder work.

The capability list is specific. Both run tool and API calls in the background while dialogue continues. Both handle near real-time visual input and 97 languages, with mid-conversation switching. Both emit early verbal cues — a "Let me check that…" while the call is in flight — which is the small trick that makes latency survivable. Extended Thinking goes further and reasons while it speaks, narrating progress on multi-step tasks.

The scores back it up. Artificial Analysis put Extended Thinking (High) at #1 on its Speech to Speech Index at 82.6, and #1 on its τ-Voice implementation at 68.6%. Google also cites 35.1% on Sierra's τ-Voice-banking benchmark and 97.7% on Big Bench Audio.

Availability: Gemini API and AI Studio for developers, over stateful WebSocket connections via the Live API. Private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience coming. Consumers get it in Search Live, Gemini Live, and Workspace for Pro and Ultra subscribers. Google published no pricing in the announcement — a detail worth holding onto, given it doubled Gemini 3.8 Flash pricing effective January 1.

Why background tool calls matter for your business

Dead air is the whole problem. A voice agent that has to look up an order, check stock, or hit your CRM either goes silent for four seconds or reads a canned holding line. Callers hang up on both. Running the tool call under the conversation removes the structural reason your front desk bot feels broken.

τ-Voice-banking at 35.1% is the number to read, not 82.6. Quality indexes measure how it sounds. The agentic benchmarks measure whether it finishes the task. A third of banking-style transactions completed is a real result and also a clear ceiling — this is an assistant for your front line, not a replacement for it. Build the escalation path before you build the agent.

Do not wire your logic into one vendor's WebSocket. Stateful Live API sessions are convenient and sticky. Keep your prompts, tool definitions, and transcripts in your own system so a price change or a better model is a config swap, not a rebuild. We have made this argument about pinning voice model aliases and it applies harder here, because a voice stack touches your phone number, your CRM, and your customer's patience at once.

No pricing means no budget. Do not move a production phone line onto a model whose per-minute cost is unpublished. Prototype now, commit when the rate card exists.

Sources: Google announcement, Artificial Analysis.