Google has replaced Search Live’s voice model with a new one. On Sep 15 2026 the company unveiled Gemini 3. 8 Live and Gemini 3. 8 Live Extended Thinking, two live dialogue models that are already rolling out via the Gemini API, Google AI Studio, Gemini Enterprise, Gemini Live, Google Workspace, and Search Live.
Tom Ouyang, a principal engineer at Google, and Malini Jaganathan, a member of technical staff writing for the Gemini Audio Team, made the announcement. The two models were described as “our most advanced live dialogue models yet, ” that have the capabilities for natural conversation and voice driven task execution.
As reported by Search Engine Land, this change was already live on Google Search. For search marketers, which model is running where is the key thing. Search Live, the voice feature inside Google Search, runs on Gemini 3.8 Live, while the heavier Extended Thinking model went to Gemini Live and parts of Workspace.
Two Models With Different Jobs
Google split the release along cost and capability lines.
Gemini 3. 8 Live is the volume model. Google says that it is built for high volume and cost efficiency by combining conversational ability, smooth dialogue, and visual grounding. It’s the version that handles most consumer surfaces.
Gemini 3. 8 Live Extended Thinking is the thinking (reasoning) model. It aims at high complexity problems requiring steps of reasoning and is priced because of this. Google is positioning these two as the base layer for production voice agents that developers and enterprises will use.
The Benchmark Claims
Google published numbers for Extended Thinking. The model led Artificial Analysis’ Speech to Speech Quality Index with a score of 82.6. On agentic task completion it posted 68.6% on the τ-Voice benchmark and 35.1% on Sierra’s τ-Voice-banking test.
Gemini 3.8 Live, the lighter model, placed second in the Speech Agent Arena. This was based on a people’s preference ranking rather than pure capability rating. That distinction is worth holding onto. The Arena’s ranking shows which model the people preferred to talk to in the direct comparison. That means that, in a sense, the performance there says more about how conversational the model was than about whether it did the job right or not. Google used both kinds of measurements back-to-back in the same announcement and they clearly have different purposes.
In addition, Google mentioned ServiceNow’s EVA-Bench, an evaluation tool for voice assistants. They claimed that their models achieve a good trade-off between factual precision and conversational style through complex sequences. The company did not put a single result for EVA-Bench. They stated the test was run on the Live API within Gemini Enterprise Agent Platform.
What the Models Actually Do
Three capabilities account for most of the release.
Gemini 3.8 Live processes visual input in near real time, so a user can point a camera at something and ask about it while the conversation continues. By looking at a subject through a camera one can interactively query it without the conversation breaking. Google demo videos have shown it taking a new worker through their onboarding step by using only what appears on the screen and also playing chess using a live stream of the game from a camera.
The model also detects and switches between 97 supported languages mid conversation, without the user selecting anything.
Third, it runs tool calls and API requests in the background while still talking. Instead of going silent during a lookup, it acknowledges the request and keeps the dialogue going until the task finishes. That removes the dead air that made earlier voice assistants feel like phone menus, where every database query produced a pause with no explanation.
Extended Thinking adds simultaneous reasoning and speech. Google says the model uses short verbal cues such as “Let me check that” to acknowledge a prompt, and then it will explain the different stages of background work through which it is passing as it narrates them. The demonstrations show it converting handdrawn sketches into actual functioning code with React through voice instructions, multi-step booking with asynchronous function calls coordination, and writing business plans and marketing kits by voice only.
Search Live Gets the Cheaper Option
Search Live is where most people will hit these models without knowing it. Search Live will be the interface for that. Google mentions that the Gemini 3. 8 Live model can be used for step by step troubleshooting help in Search Live, such a situation when a user shows a product to their camera and asks what to do next.
Google went for the cost-optimized model for its primary consumer-facing service Search Live. Search Live handles enormous query volume, and latency and unit cost both scale with it. Extended Thinking sits behind subscription products, where the economics are different.
This segmentation also informs us of the capabilities of Search Live. It gets visual grounding, the 97 language handling, and background tool execution. It does not get the simultaneous reasoning and narration that Google demonstrated with Extended Thinking, so long multi step tasks are not the target here. The main use scenario for Search Live is when users have an object in front of the camera and they need guidance on how to interact with that object.
Where Extended Thinking Landed
For the general consumers, this feature of Extended Thinking is found in Gemini Live. In Workspace, it reached Docs for Google AI Pro and Ultra subscribers, and Gmail and Keep for all Google AI subscribers.
The two models can be used both for Gemini Live API as well as Google AI Studio. Private preview for enterprise access will be launched in parallel with Gemini Enterprise, and Gemini Enterprise for Customer Experience is listed as coming soon.
The Partner List
Google named developer platforms that already support the Live API, including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents. Picking on the customer side, Google identified Salesforce, Genspark, and Lumeris as customers already taking advantage of the feature.
Google cited latency, fluidity, and tool calling as the reasons Salesforce, Genspark, and Lumeris signed on. In fact, Google included testimonial cards of ServiceNow LiveKit Agora Lenskart 11Sight, Equal AI, Ambr AI, and Casuu from the customer and partner side in the release. These partners and customers span different sectors like customer service, retail, and health care.
Release Pace and Watermarking
The Gemini 3. 8 has been launched twice in a span of 2 weeks. Initially, Google introduced Gemini 3. 8 Flash, Flash Cyber variants and then audio models. The text branch of the Gemini series has had multiple releases this year. This release moves the same version number onto the audio stack.
The audio that these models create will have the watermark called SynthID included in the output. Google publishes its safety notes for the release in the Gemini 3.8 Audio model card.
What Search Marketers Should Watch
Voice queries inside Search Live do not produce a results page in the usual sense. The person speaks, the system answers, and there is no ranking competition between ten blue links. Tracking how often a brand or product surfaces in those answers is harder than tracking position on a SERP, and Google has not opened reporting for it. Search Console just doesn’t report on the Search Live feature, and this announcement gives no clues that is going to change.
The visual grounding part is probably something to experiment with first. If a user can hold up a product and get troubleshooting steps back, support content and product documentation become retrieval targets in a way they were not before. Well organized, structured and properly marked-up support content is what the model can work and pull from correctly. Those support pages that hardly explain anything yet manage to attract long-tail search traffic will be of little help in this case since the model will look for the specific steps, not the webpage that is optimized to rank for the issue.
For people who operate websites in more than one language, language handling is worth another glance. If the model is able to switch languages during the conversation, a user can actually begin the query in one language and conclude in another, and this is not straightforward to handle for the content which is usually created separately for different markets.
FAQs
Q 1: What is Gemini 3.8 Live?
Gemini 3.8 Live which is a live dialogue model announced by Google on 15th September 2026. It is designed to support scale and cost with voice conversation, visual input, and background tool calls. it processes voice conversation, visual input, background tool calls, etc.
Q 2: How is Extended Thinking different?
Gemini 3.8 Live Extended Thinking higher capability is developed for multi-factor interaction, multi step reasoning that calls for more complex and high-level cognitive processes, in which “thinking” and “talking” take place simultaneously.
Q 3: Which model powers Search Live?
Search Live is based on Gemini 3.8 Live, not on Extended Thinking. Google directed the less costly model to the surface of highest volume consumer.
Q 4: How many languages does it support?
Gemini 3.8 Live is compatible with 97 languages. It automatically recognizes and switches languages in the middle of a conversation.
Q 5: What did it score on benchmarks?
Extended Thinking was the best Artificial Analysis ranked Speech to Speech Quality Index with a score of 82.6. It scored 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking benchmark, and 97.7% on Big Bench Audio.
Q 6: Can developers use these models now?
Yes. Both are available through the Gemini API and Google AI Studio effective September 15, 2026.
Q 7: What about enterprise access?
Enterprise access is in private preview via Gemini Enterprise. Per Google, Gemini Enterprise for Customer Experience is coming soon.
Q 8: Where does Extended Thinking appear in Workspace?
Has been rolled out to Docs for Google AI Pro and Ultra subscribers, Gmail and Keep for all Google AI subscribers. Workspace business customer access plans to be available soon.
Q 9: Is the generated audio watermarked?
Yes. These models’ output all contain a SynthID watermark encoded in the produced audio so AI generated speech remains identifiable.
Q 10: Does this change how sites rank in Search?
Not directly. Voice answers in Search Live do not produce a standard results page, so there is no new ranking system attached to this release, and Google has not added reporting for it.
