Google Adds Gemini 3.8 Live and an Extended Thinking Voice Model

Google has released two new real-time models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, aimed at voice agents that hold a conversation while doing work in the background. The company positions the first as the cost-efficient option for conversational AI and visual grounding, and the second as the model for high-complexity, multi-step tasks. Both began rolling out on September 15, 2026 across the Gemini API, Google AI Studio, a private preview in Gemini Enterprise, and a set of consumer surfaces.

What was announced

The split between the two models is the most important detail. Gemini 3.8 Live is the everyday tier: fast, cheaper, meant for conversational turn-taking and for understanding what a camera is pointed at. Gemini 3.8 Live Extended Thinking sits above it for work that needs several reasoning steps before the model can answer.

Both models handle near real-time visual input, which is the piece that separates this generation of voice models from the text-to-speech pipelines that preceded them. The model is not transcribing audio, thinking, and then reading a reply aloud; it is taking speech and video as native input and producing speech as native output.

Google also says the models detect the language being spoken and switch between 97 supported languages mid-conversation, without the user having to declare a language up front or restart a session. For anything customer-facing in a multilingual market, that removes a routing step that was previously handled by a separate language-ID model.

The third capability worth flagging is background tool execution. According to Google, the model can call tools and APIs without interrupting the conversation. Extended Thinking goes further: it reasons and speaks at the same time, and Google describes it emitting early verbal cues such as "Let me check that…" while the reasoning is still in flight.

Why the speaking-while-thinking part matters

Latency has been the hard constraint on voice agents. A model that needs two seconds of silence to decide what to do reads as broken to a caller, even when the eventual answer is correct. Every production voice system has worked around this with filler audio, hold music, or a scripted acknowledgement played by the orchestration layer while the model catches up.

Moving that behavior into the model itself is a meaningful simplification. If the model knows it is about to take three seconds and says so in its own voice, the orchestration layer stops having to guess. Whether it holds up under real call volumes is a separate question, and Google's post does not publish latency figures.

The background tool execution claim has the same shape. Today most voice agents freeze when they hit a function call, because the conversational loop and the tool loop are the same loop. Decoupling them is the right architecture. Teams evaluating this should test it against the ugly cases: a tool that times out, a tool that returns an error, a user who interrupts while a call is in flight.

The benchmark numbers Google published

Google reports that Gemini 3.8 Live Extended Thinking scores 82.6 on the Speech to Speech Quality Index and ranks first overall. On agentic task completion it cites 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking. For audio reasoning, Google reports 97.7% on Big Bench Audio. On Speech Agent Arena, Gemini 3.8 Live is listed at #2. Google also says the models push the Pareto frontier on ServiceNow's EVA-Bench for complex workflows.

Two of those numbers are more informative than the rest. The 97.7% on Big Bench Audio is high enough that the benchmark is close to saturated, which makes it weak as a discriminator between frontier models. The 35.1% on τ-Voice-banking is the opposite: a domain-specific agentic benchmark where the model completes roughly a third of tasks. That gap between a near-perfect audio reasoning score and a one-in-three completion rate on banking workflows is the honest summary of where voice agents currently sit. Understanding speech is largely solved. Reliably finishing a multi-step task on a customer's behalf is not.

Google does not publish a comparison table against competing models in the announcement, so the rankings have to be taken at the company's framing.

Availability and who gets what

Gemini 3.8 Live is rolling out to developers through the Gemini API and Google AI Studio, to enterprises through a private preview in Gemini Enterprise, and to everyone through Search Live.

Gemini 3.8 Live Extended Thinking has a wider consumer footprint. Developers get it in the Gemini API and AI Studio. Enterprises get the same private preview, with Google saying it is coming to Gemini Enterprise for Customer Experience and to Google Workspace business customers. On the consumer side it reaches Gemini Live, Google AI Pro and Ultra subscribers in Workspace, and all AI subscribers in Gmail and Keep.

Google says all AI-generated audio carries SynthID watermarking.

Pricing is not disclosed in the announcement, which is the largest practical gap. Real-time audio models are billed differently from text models, often per second of audio in and out, and the cost profile of a model that reasons while speaking is not obvious. A team cannot size a voice deployment without it.

Ecosystem and open questions

Google lists integrations with Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents, which covers most of the real-time transport and agent-orchestration stack teams already use. That matters more than it sounds: voice agents are rarely built directly against a model API, and support in LiveKit or Pipecat is what determines whether a prototype ships.

The company also cites testimonials from Salesforce, Genspark, Lumeris, ServiceNow, Lenskart, and Agora. These are partner statements published alongside a launch, not independent evaluations, and should be read that way.

Several things the announcement leaves open are worth naming. There is no stated rate limit or concurrency guidance for the API. There is no detail on how the two models are meant to be used together, though the obvious pattern is routing simple turns to 3.8 Live and escalating to Extended Thinking. And the 97-language claim covers detection and switching, not necessarily equal quality across all 97.

What to watch

• Whether Google publishes per-minute or per-token pricing for both models, and how Extended Thinking's reasoning time is billed during a live call.
• Real-world latency under concurrency, particularly whether the speaking-while-thinking behavior degrades when many sessions run at once.
• How the private preview in Gemini Enterprise converts into general availability for Gemini Enterprise for Customer Experience and Workspace business customers.
• Independent results on agentic voice benchmarks, given the wide spread between the audio reasoning score and the τ-Voice-banking completion rate.
• Whether the background tool execution path handles failures and interruptions cleanly enough for production support workloads.