--- title: "The pause is the product" date: 2026-09-20 url: https://remiam.co.uk/notes/the-pause-is-the-product tags: [AI, Voice, Installations] read_time_minutes: 5 description: "Remiam's read on Gemini 3.8 Live and Live Extended Thinking, released September 2026: why the gap between a spoken question and its answer is a design decision for any voice system in a venue, a kiosk or a service line." --- # The pause is the product *Published 2026-09-20 · 5 min read · by Liam (Remiam)* Google has shipped a voice model that talks while it thinks. The benchmark gain is real. The more useful lesson is that the silence between a question and an answer is now something you design. On 15 September Google released two voice models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both take speech in and give speech back, call tools in the background, and handle 97 languages with switching mid-conversation. They went generally available through the Gemini API the same day. The headline number is the gap between the two. On the τ-Voice agentic benchmark the standard model scores 30.1 per cent and the Extended Thinking variant 68.6 per cent. Same voice, same interface, more than twice as good at finishing a multi-step task, because it is allowed to reason before it commits to an answer. ## Thinking out loud, literally The interesting part is how it spends that reasoning time. A text model can think for eight seconds and nobody minds, there is a spinner. A voice model that goes silent for eight seconds sounds broken. So Extended Thinking fills the gap. It gives an early cue along the lines of "let me check that", streams conversational filler while it works, and speaks progress updates on longer tasks. That is not a model feature so much as a front-of-house decision that has been made for you. Anyone who has worked a counter knows it already. A customer will wait a surprisingly long time if they can see you are dealing with it, and almost no time at all if they cannot. > On a screen, latency is a number. In a room, it is a silence, and people read silence as failure. ## Where this matters for the systems we build Most of our voice work is not a phone assistant. It is a kiosk in a busy venue, an installation people walk up to, a check-in point with a queue behind it. In those settings the pause is the whole experience. Three things follow. - Decide what the silence sounds like before you pick a model. A filler phrase, a sound, a light on the unit, a line of text on the screen. If you leave it to the model you get its default, and its default was tuned for a phone call, not a noisy hall. - Match the thinking level to the question. Extended Thinking offers low, medium and high reasoning settings. "Where is the cloakroom" does not need deliberation. "Can I move my booking to Thursday and keep the discount" does. One model per install is the wrong unit, the routing is per question. - Budget for the minute, not the token. Google prices the standard model at about $0.005 a minute of audio in and $0.018 a minute out. An installation that talks all day is a line on the running costs, not a rounding error, and reasoning tokens on Extended Thinking bill on top. ## The number to read twice The same release reports 35.1 per cent on τ³-Banking, a benchmark built around regulated, rule-heavy service conversations. That is a leading score and it still means most of those tasks did not complete correctly. It is the useful reality check. A voice model that sounds calm and competent while it works is a very different thing from one that gets the refund right. So our advice has not changed much, only sharpened. Let voice handle the conversation, and let a system you can test handle the decision. The model can say "let me check that" as warmly as it likes. What it checks against should be your booking rules, your stock, your policy, called as a tool and logged, not reasoned from scratch in the middle of a sentence. There is also a quieter engineering note for anyone migrating. Extended Thinking reports progress through a new interaction status field rather than the old turn-complete signal, so the state handling in an existing voice app needs rewriting, not just a model name swapped. Plan for that before promising a client a same-day upgrade. Voice has been close to good enough for a while. What changed this week is that the gap it used to leave is now something you are expected to fill on purpose. ## References 1. [SiliconANGLE, Google's new speech model Gemini 3.8 Live supports real-time reasoning, 15 September 2026](https://siliconangle.com/2026/09/15/googles-new-speech-model-gemini-3-8-live-supports-real-time-reasoning/) 2. [DataCamp, Gemini 3.8 Live: Features, Benchmarks, Pricing, Access, September 2026](https://www.datacamp.com/blog/gemini-3-8-live) 3. [Digital Applied, Gemini 3.8 Live: Should a Voice Agent Think While Talking?, September 2026](https://www.digitalapplied.com/blog/gemini-3-8-live-extended-thinking-voice-agent-choice)