Gemini 3.8 Live Can Think While It Talks — and Its Model Card Adds Some Caveats

The takeaway: Google has released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two models built for spoken conversation, which it calls "our most advanced live dialogue models yet". Google's pitch centres on how they talk. They can keep talking while they think, carry on the conversation while a tool call runs in the background, and switch languages without breaking stride. Google says both come with "Major upgrades in intelligence and parallel reasoning". The model card is worth reading alongside the launch post for its caveats.

Two models, two jobs

Google has split the release by workload:

• Gemini 3.8 Live is "Built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding." It's the high-volume model for voice agents that need to be quick and cheap.
• Gemini 3.8 Live Extended Thinking is "Built for high-complexity tasks, with increased intelligence and multi-step reasoning." It's for longer, messier jobs where the model has to work something out while it's still on the line.

What's actually new: the conversation doesn't stop

Voice assistants have tended to have one awkward trait: when they have to think or fetch something, the conversation stalls. Google's pitch is that this goes away.

• Talking while thinking. Google says Extended Thinking "reasons and speaks simultaneously", using early verbal cues like "Let me check that…" and narrating its progress through multi-step tasks rather than going silent.
• Background tools. Gemini 3.8 Live "executes tools and API calls in the background while continuing the conversation", so it can acknowledge a request and keep chatting while the task finishes.
• Languages. It detects and switches between 97 supported languages mid-conversation, according to Google.
• Vision. Gemini 3.8 Live takes in visual input in near real time. Google's demos include it playing chess from what it sees, and Extended Thinking turning rough sketches and spoken feedback into working React components.

The benchmark claims

All of these figures come from Google's own announcement:

• Extended Thinking takes first place overall on Artificial Analysis' Speech to Speech Quality Index, with a score of 82.6. It scores 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark for agentic tasks, and 97.7% on Big Bench Audio.
• Gemini 3.8 Live placed second in the Speech Agent Arena, which Google cites as evidence of high user preference.

Google hasn't published prices in the post. It says Extended Thinking keeps a "highly competitive price point compared to other frontier models" and that the standard model stays cost-effective.

What the model card says

The model card says both models are based on Gemini 3 Pro, and adds some useful caveats:

• Knowledge cutoff. Google states: "The knowledge cutoff date is January 2025." That's a long gap for a model launched in September 2026. Unless it's connected to search or tools, anything it tells you about recent events should be treated with care.
• Context. Both models take audio, images, video and text, with a context window of up to 128K tokens and up to 64K tokens of output.
• Known limitations. The card warns the models may show general foundation-model failings "such as hallucinations", and there may be "occasional slowness or timeout issues". That second one matters more for a voice product than for a chatbot.
• Frontier safety. Under its Frontier Safety Framework, Google's assessment found that neither model has meaningful new capabilities or material increases in performance compared with Gemini 3.7 Flash, which had already been tested. On that basis, Google says the new models are unlikely to reach any of its tracked or critical capability levels.

That last point is about safety thresholds, not everyday usefulness, so it doesn't contradict Google's claims of better intelligence. It does mean that, by Google's own assessment, the release isn't a step change on the safety risks it tracks.

Where you can use it

Both models are rolling out from launch:

• Developers: the Gemini API and Google AI Studio, via the Live API.
• Enterprises: private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience "coming soon" (and, for Extended Thinking, Google Workspace business customers).
• Everyone: the standard model powers Search Live. Extended Thinking is in Gemini Live, and in Workspace for Google AI subscribers. Docs support is for AI Pro and Ultra subscribers, while Gmail and Keep are available to all Google AI subscribers.

Google says all audio from its AI products is watermarked with SynthID, an imperceptible marker built into the audio so AI-generated speech can be detected.

Why it matters

In our view, much of what has made voice agents feel like phone trees is the awkward pauses: the silence while the model thinks, and the conversation stalling while it books something or looks it up. If Google has really fixed that, it matters for customer service lines and hands-free work, whatever the benchmark tables say. Just check the model's facts: its knowledge stops at January 2025.