Search Live Now Runs on Gemini 3.8 Live — and It's Available to Everyone

Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September 2026, in a post by Alisa Fortin and Thor Schaeff of Google DeepMind. Alongside them came Gemini 3.5 Transcribe, a speech-to-text model.

The line that matters for search: Gemini 3.8 Live is available to everyone in Search Live. Not a Labs experiment, not a subscriber tier — the general model behind Google's conversational voice search has been upgraded for all users.

Extended Thinking, the heavier reasoning variant, goes elsewhere: Gemini Live, Docs for AI Pro and Ultra subscribers, and Gmail and Keep for Google AI subscribers.

What changed in the model

These are native speech-to-speech models — audio in, audio out, without a transcribe-then-generate-then-synthesise pipeline in between. That architecture is why the conversational behaviour is different rather than just faster.

The capabilities Google lists:

• Asynchronous function calling — the model can trigger and run background operations while it is still speaking, instead of going silent during a lookup.
• Visual context grounding — live video input feeds the dialogue, so the conversation can be about what the camera is pointed at.
• 97+ languages, with switching supported mid-conversation.
• Alphanumeric precision for codes, reference numbers and technical strings — historically where voice models fall apart.
• Incremental content updates, merging spoken audio with structured data as it arrives.

Extended Thinking adds parallel reasoning and narrates its own progress with verbal cues — the "let me check that" pattern — while background work completes. Google says it ranks first on Artificial Analysis' speech-to-speech leaderboard.

Pricing on the API is $0.005 per minute of audio input and $0.018 per minute of output.

Why an audio model release is a search story

It's tempting to file this under AI model news and move on. Three reasons not to.

Asynchronous function calling changes the shape of a voice query. Previously, a conversational assistant that needed to fetch something had to stop talking to do it, which creates a hard pressure toward short answers and few lookups. Remove that constraint and a single spoken query can support several retrievals without the conversation stalling. More retrievals per query means more opportunities for a source to be pulled in — and more competition within each one.

Visual grounding expands what counts as a query. A user pointing a camera at a product, a fault, a landmark or a page of text and asking about it is issuing a search that has no typed equivalent and no keyword. There is no query string to rank for. What gets surfaced is decided by whatever the model retrieves about the entity it has recognised — which puts the emphasis on your entity being unambiguous and well described across the web, rather than on any page matching a phrase.

97 languages with mid-conversation switching is a genuine shift for anyone doing international SEO. The assumption that a user stays in one language for one session is now wrong by design.

The measurement problem this makes worse

Only yesterday we covered John Mueller conceding that Search Console's generative AI reporting doesn't work properly — that impressions count links nobody saw and content behind "Show More" isn't counted at all.

Voice makes every part of that harder. There is no viewport, no scroll depth, no visible list of ten results, and frequently no click at all. A spoken answer that draws on your content and satisfies the user completely may generate no measurable signal of any kind.

So this upgrade lands in a gap: Google is shipping meaningful capability into conversational search faster than it is shipping any way to see what that capability does to your traffic. Plan on that gap persisting.

What to do

Don't rebuild anything for voice. There is no separate voice index and no voice-specific markup to add. The retrieval layer is the same one you're already optimising for.

Get your entity right. Visual and conversational queries resolve to entities rather than keywords. Consistent naming, accurate structured data and correct sameAs relationships are what let a model identify what it's being shown and find you as a source on it.

Write answers that survive being read aloud. A claim that depends on a table, a screenshot or surrounding layout to make sense doesn't survive the trip through a speech-to-speech model. Self-contained, plainly stated sentences do.

Check your factual specifics. Alphanumeric precision being called out as an improvement is a reminder that prices, model numbers, opening hours and reference codes now get spoken back to users verbatim. Wrong ones on your site become wrong ones in someone's ear.

Expect follow-up queries, not single questions. Conversation that doesn't stall is conversation that continues. Content that anticipates the second and third question in a sequence has an advantage over content that answers only the first.

Sources

• Google: Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe
• Google AI for Developers: Live API documentation