Gemini 3.8 Live: check the record, not the chat

The camera dies when the phone locks. The voice can keep talking.

That is Google’s own Search Live behavior, not a metaphor. Spoken follow-ups continue; live camera context does not. If you pointed the lens at a label and then locked the phone, the assistant has lost that view. A useful answer should notice. A bad one keeps talking as if the label is still there.

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking were announced September 15 (post updated September 17). Google says they can use visual context and keep a conversation moving while work happens in the background. Published demos include coordinating bookings and turning sketches plus spoken feedback into React components. Those are company demonstrations — not Daybreak tests.[1]

You shouldn’t need a perfect prompt. You should be able to talk, point, interrupt — and still get the booking right. The hard part isn’t the chat. It’s whether Thursday at 2 becomes Thursday at 3 when you change your mind mid-call.

Sounding like a collaborator and completing the right action are different jobs. Independent scores make that split visible.

Where you can actually use it

Google says standard 3.8 Live rolls out through Search Live and developer tools. Extended Thinking rolls out through Gemini Live and developer tools, with different Workspace subscription rules. Enterprise pieces include private previews and “coming soon.” An announcement is not a guarantee that your account has every demo.[1]

Concrete start: Search Live in the Google app on Android and iOS, where AI Mode is available. Spoken follow-ups, web links, optional camera sharing. Camera stops when the app backgrounds or the phone locks — even if voice continues.[2]

Model capability ≠ finished product. A developer can wire a Live model to a scheduling API. That does not prove your Search Live session can book the same table. The app has to expose the tool, the access, and the rules.

What happens behind the voice

Google’s Live API docs describe a continuous connection for sound, images, and text with spoken replies. Interruptions, transcripts, and tool use are listed features.[4]

The part that gets something done is the tool. The app defines functions the model may request, then reports what happened. “Turn on a light” is a function call plus a result — speech alone is not the action.[5]

Background functions can keep the conversation open while work runs. Results may interrupt, wait for a quiet moment, or become later context. Support varies by model; some tables still describe older Live versions. Principle: talk and external work do not have to take turns owning the whole call.[5]

Walk a booking change. You ask for Thursday afternoon. While the scheduler checks, you say you cannot arrive before three. A useful assistant updates the request, finds a slot, and confirms the record. If the first request already went through, it says so and fixes the conflict.

Filling silence is easy to notice. Keeping the current request, search vs reservation, and pending vs done is what decides whether you have an appointment.

The stronger thinker is not better at every part

Artificial Analysis (checked September 21): Gemini 3.8 Live Extended Thinking (High) completes 68.6% of simulated customer-service tasks on τ-Voice. Standard 3.8 Live scores 30.1%. On a separate conversational-dynamics score, the order flips — standard Live 96.1%, Extended Thinking 91.9%. Two tests. Not a combined real-world success rate. Not a prediction for your next call.[6]

Task eval: agent gets tools and policies in simulated airline / retail / telecom settings; score checks final database state. Conversational measure: pausing, interruptions, continuing after a brief “uh-huh.” Sounding good and leaving the database correct are different jobs.[7]

A benchmark is not a live call center. The original τ-Voice work uses simulated callers. Its March preprint found noise and accent variety hurt then-tested voice agents. That paper does not measure Gemini 3.8. It does explain why a clean demo undersells the hard parts.[8]

A big jump on a hard task test matters. It still does not tell you whether a given airline wired working tools, whether booking is available in your product, or what happens when the call drops.

The competition is also learning to listen

September 18: SpaceXAI released Grok Voice Transcribe 2.0 — speech-to-text for hard audio (overlap, phone lines, accents, spoken account numbers). Different product from Gemini’s conversational models. A perfect transcript does not finish a task.[9]

Fail early and the rest of the chain is wrong. Miss a digit, miss a correction, and the model solves the wrong problem. Nail the transcript and still send the wrong API call, and you fail later. Hearing → interpreting → acting → confirming. All of it.

Less repetition is a better user standard than “sounds human.” Carry the problem through the exchange. Accept a correction once. Keep it when the task hits another service.

Companies get a better scorecard than call length alone. Was the problem resolved? Did the person understand what changed? Was another call needed? Those are proposed usefulness checks — not outcomes this release proved.

What to check after the demo

Google’s model card: Gemini 3.8 Audio can hallucinate; it can hit slow responses or timeouts. Audio, images, video, and text in ≠ verified observation out. Knowledge cutoff listed as January 2025 — built-in knowledge is not the same as what a live retrieval brought in this session.[10]

Practical trial: pick an outcome you can inspect. Show a non-sensitive object; ask something checkable. If the app supports actions, open the document, appointment, or record. Mid-call, change your mind once. See whether the correction survives into the final state.

Taste call: judge the release on whether the record matches the last thing you said — not on whether the voice filled the silence.

Search Live’s camera dies on lock while voice continues. That single product fact already separates “still talking” from “still seeing.”

Continuous voice, camera context, background tools — closer together. The scores still force a pick: task completion or conversational gloss. Not one universal winner.

Wait for a booking or ticket you can open yourself. If the assistant says “done” and the record disagrees, that is the story — not the demo reel.