Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026. The two real-time audio models are designed to keep a voice conversation moving while they process visual input, call tools and, in the Extended Thinking version, work through complex multi-step tasks.
Two models, two different jobs
Gemini 3.8 Live is aimed at scalable, cost-efficient dialogue with low latency. It can process visual input in near real time and run tool or API calls while the conversation continues.
Gemini 3.8 Live Extended Thinking targets more demanding workflows. Google says it can reason and speak simultaneously, using verbal acknowledgments and progress narration while it works through a longer task.
| Capability | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
| Primary focus | Scalable, low-latency live dialogue | Complex, multi-step reasoning |
| Conversation behavior | Fluid dialogue with visual grounding | Conversation continues while reasoning and speaking |
| Tool and API calls | Can run while the conversation continues | Can run in the background while reasoning and speaking continue |
| Visual input | Near-real-time processing | Supported during live multimodal interaction |
| Best fit | High-volume voice interactions | Workflows that benefit from deeper reasoning and progress narration |
A small developer comparison used the same prompt, instructions, tool and session code across six runs. Gemini 3.8 Live completed the drawing-tool call in two of three sessions, while Extended Thinking completed it in three of three and produced four to eight audible progress segments before each call. That sample illustrates the intended behavioral difference; it is not a general reliability benchmark.
The important change happens while the conversation continues
The Gemini Live API demonstration shows asynchronous function calling, proactive audio and client-context injection through send_client_content. It also displays tool calls and their results in real time.
In practical terms, a voice agent can keep the dialogue active while a background operation runs. The interaction is not limited to the familiar pattern of speaking, waiting in silence and then receiving a final answer. Extended Thinking applies the same idea to reasoning: Google’s model is designed to continue the exchange while it works through a more involved request.
The demonstration includes an order lookup running in the background and a comparison in which Extended Thinking generates an SVG of a pelican riding a bicycle while maintaining the conversation. These are demonstrations of the Live API and model behavior in those sessions, rather than a guarantee that every application will behave identically.
Languages, visual input and reported performance
Google says the Live models can automatically detect and switch between 97 supported languages during a conversation. The models also accept visual input, allowing a voice agent to respond to what it sees as well as what it hears.
Google reports that Gemini 3.8 Live Extended Thinking scored 82.6 on Artificial Analysis’ Speech to Speech Quality Index, 68.6% on τ-Voice agentic task completion, 35.1% on Sierra’s τ-Voice banking benchmark and 97.7% on Big Bench Audio. These figures are reported results for the Extended Thinking model and use different evaluation measures; they are not one universal performance score.
Google also says generated audio carries a SynthID watermark. For developers building a voice agent, the more consequential question is how the system handles interruptions, noisy audio, accents, confirmation of names and numbers, recovery after a dropped call and visible records of external actions. Natural-sounding speech is only one part of a useful agent.
Developer access and consumer rollout
Google began rollout through the Gemini API and Google AI Studio. The announcement also lists Search Live, Gemini Live, selected Gemini experiences, Workspace surfaces such as Docs, Gmail and Keep, and a private preview for Gemini Enterprise, with access varying by model and product.
For US users, those channels describe the rollout Google announced; they do not guarantee that every account has access to every model or experience. The Gemini Live API is the developer-facing route for building applications around the real-time audio features.
What the launch changes for voice agents
The key shift is concurrency. Gemini 3.8 Live is designed to handle dialogue, visual context and tool execution without stopping the exchange, while Gemini 3.8 Live Extended Thinking adds a deeper reasoning mode for multi-step work.
That makes the models relevant to applications such as assistants that retrieve information, interpret a camera view or interact with external tools during a spoken session. The useful test is whether an agent can preserve context and clearly communicate its actions while doing that work—not simply whether it can produce a convincing voice.
