News · 2026-09-16

Google ships Gemini 3.8 Live and a separate Extended Thinking voice model

Google has begun rolling out Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking as two separate realtime voice-model endpoints. The ordinary model prioritizes low-latency conversation, visual grounding and non-blocking tools; Extended Thinking is designed to reason and call tools in the background while it keeps speaking.

Key facts

• Google says Gemini 3.8 Live can switch automatically among 97 languages and process visual input in near real time.
• Extended Thinking High scores 82.6 on Artificial Analysis’s current composite speech-to-speech index; ordinary Live scores 76.0.
• Artificial Analysis’s blind Speech Agent Arena currently places ordinary Live second, with 1,083 Elo from 482 samples and a 93.2 percent task-success rate.
• The primary source is Google’s launch announcement, supported by the Live API model documentation.

The distinction matters because “voice AI” hides two different user experiences. Ordinary gemini-3.8-live is the fast conversational endpoint. It supports streaming audio, multimodal input and asynchronous function calls by default. The client can acknowledge a request, continue talking, and receive an API result later rather than holding the conversational turn hostage to a database query. The previous 3.1 Live documentation described synchronous calling, so this is a real change in agent orchestration rather than a cosmetic model rename.

gemini-3.8-live-extended-thinking is a different endpoint, not a hidden reasoning toggle. Google says it is for multi-step voice work and can narrate progress while reasoning or tools operate in the background. The thinking documentation adds an important operational wrinkle: turnComplete no longer means the system is idle, so a client must follow interaction status. The analogy is a receptionist who can keep the caller informed while a specialist looks through records; useful, but more complicated than a single request-response call.

The best public evaluation signal is nuanced. Artificial Analysis’s speech-to-speech table lists Extended Thinking High at 82.6, including 68.6 percent agentic performance and 97.7 percent on Big Bench Audio. But ordinary Live does better on some user-facing measures, including the current blind arena. Google’s launch post says the models reach an EVA-Bench Pareto frontier, yet the most defensible claim is benchmark-specific: Extended leads this composite index; ordinary Live is the stronger current blind-conversation product. It is not sound to flatten those results into “Google has the best voice model.”

Availability is staged rather than universal. Google says developers can access the models through the Gemini API and AI Studio; enterprise access is private preview. Consumer distribution differs by endpoint: ordinary Live is in Search Live, while Extended Thinking is in Gemini Live and selected Workspace surfaces. The release should not be rewritten as blanket availability across every Gemini consumer surface. Google’s own model page for Extended Thinking is the implementation reference.

There is a separate warning for agent builders. A supplied PrimeTime transcript records one engineer saying Gemini 3.8 Flash in Cursor reread a file repeatedly for about 40 minutes, consuming “330 million tokens” and “$118” without fixing a small color bug. That is a concrete anecdote, not a reproducible result or a diagnosis of whether Cursor’s loop or the model caused the behavior. It nevertheless captures the product lesson: a strong voice score does not supply spend limits, loop detection, cancellation or tool-side safety. Voice agents should cap turns and tool retries just as coding agents do.

Google’s launch quote is its product claim that the system delivers “more natural, capable and helpful” real-time interactions. The measurable anchor is the 97-language capability, not an assertion of general intelligence. The honest caveat is that the Live API remains preview, documentation has shown version drift in language counts, and the benchmark rankings will move. A developer should also treat audio interruption, stale tool results, consent to record, and degraded-network recovery as product requirements rather than model options. Still, asynchronous work during continuous speech is a meaningful shipping step: it turns a voice assistant from a talk-then-wait interface into an application that can converse while it acts.

Primary source, verified: read the paper →

Ground Truth. (2026, September 16). Google ships Gemini 3.8 Live and a separate Extended Thinking voice model. Ground Truth. https://groundtruth.day/news/google-gemini-3-8-live-extended-thinking.html

@misc{groundtruth:google-gemini-3-8-live-extended-thinking,
  title  = {Google ships Gemini 3.8 Live and a separate Extended Thinking voice model},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/google-gemini-3-8-live-extended-thinking.html}
}

Topics: models · voice-ai · agents · tool-use · google