Gemini 3.8 Live keeps talking while it runs the tool call in the background

"Let me check that" is now a listed product feature. Google says the models use early verbal cues like it to acknowledge a prompt in a more lifelike way ^[4]. Contact centres have been running that line for decades under a plainer name. SiliconANGLE's account of Google's post describes near-real-time reasoning without a latency measurement, so how much of the wait actually shrank is not in the published material ^[21].

The scores are Google's own, and SiliconANGLE published them with the aside "if Google is to be believed" ^[8]. Inside a single model the range is wide: 97.7% on Big Bench Audio, 68.6% on T-Voice, and 35.1% on T-Voice-banking ^[7]. Banking sits 33.5 points below general T-Voice and 62.6 points below Big Bench Audio ^[19]. On that set, 64.9% of attempts were not successes ^[20]. The standard model placed second on Speech Agent Arena and first on ServiceNow's EVA-Bench ^[6].

A ten-minute call split evenly between the two sides works out at 5 x $0.005 plus 5 x $0.018, or about 11.5 cents of audio ^[18]. Extended Thinking bills reasoning tokens on top of that, along with inputs such as video and documents ^[10].

For the person who has to deploy this, the two models arrive in different places. Both are live in the Gemini API and Google AI Studio, with an enterprise private preview in Gemini Enterprise and Search Live ^[11]. Extended Thinking also reaches Workspace through Docs, Gmail and Keep for subscribers, plus the Gemini Live apps ^[12]. Existing voice stacks connect through partners including LiveKit, Pipecat, Agora, Vercel, Fishjam and Vision Agents ^[13]. The models handle 97 languages, detect the language automatically and can switch mid-conversation ^[14], and generated audio carries an invisible SynthID watermark that Google said can be used to detect misinformation ^[15].

Google AI's launch post said the models "let you speak, collaborate, and execute tasks seamlessly, meaning conversing with AI just got a lot more natural" ^[16]. Two things sort whether that holds for a given use case. The first is whether the tool call can finish while the caller keeps talking, or whether its answer gates the agent's next sentence; background execution only pays in the first case, and in the second the caller hears the filler and then waits anyway. The second is what a wrong answer at that step costs the person on the line, because a mis-stated account balance delivered fluently is a different failure from a mis-stated opening hour. For regulated, transactional work of the kind T-Voice-banking is meant to stand in for, completion rate on last week's real calls is a better guide than any index score.