Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026. The two live-dialogue models are designed for voice agents that can handle multimodal input, call tools in the background and keep the conversation moving. The split is straightforward: Gemini 3.8 Live prioritizes efficient, high-volume deployments, while Extended Thinking is aimed at complex, multi-step work.
Gemini 3.8 Live versus Extended Thinking
The models share the same basic live-dialogue foundation, but Google assigns them different jobs.
| Dimension | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
| Intended workload | Efficient, scalable voice agents and fluid dialogue | Complex workflows requiring deeper, multi-step reasoning |
| Conversational behavior | Live dialogue with background tool and API execution | Reasoning and speaking at the same time, with spoken progress during longer tasks |
| Input | Text, images, video, audio and PDFs | Text, images, video, audio and PDFs |
| Output | Text and audio | Text and audio |
| Context limits | 128,000 input tokens and 64,000 output tokens | 128,000 input tokens and 64,000 output tokens |
| Language switching | Automatically detects and transitions among 97 supported languages | Google’s stated language count for this model is not specified separately |
For a customer-service agent handling many simultaneous conversations, the standard Live model is the more direct fit according to Google’s positioning. A workflow that needs several reasoning steps—while keeping the user informed instead of falling silent—is the use case Google assigns to Extended Thinking.
What the models can do while you are speaking
Gemini 3.8 Live supports text, image, video, audio and PDF input, with text and audio output. Google also says it can process visual context in near real time, switch among 97 languages during an active conversation and execute tools or API calls while the dialogue continues.
That changes the interaction from a simple question-and-answer exchange into a live agent workflow. A voice agent can continue the spoken interaction while a background operation runs, rather than waiting for every tool call to finish before responding.
Extended Thinking adds a more explicit progress signal for demanding tasks. Google describes the model as able to reason and speak simultaneously, using phrases such as “Let me check that...” while it works through a longer operation. The goal is not merely to produce a deeper answer; it is to make the waiting period part of the conversation.
Both models have a 128,000-token input limit and a 64,000-token output limit. Google also says generated audio carries a SynthID watermark.
A voice agent needs an open conversation
The Gemini Live API uses a stateful WebSocket connection for bidirectional streaming. In practical terms, audio can travel in both directions over an ongoing connection instead of passing through a rigid sequence of speech recognition, text generation and voice synthesis steps.
The developer architecture also separates the live audio loop from the agent’s background work. A browser, backend and Gemini Live connection form the main path, while components such as the Agent, Runner, Session service and LiveRequestQueue coordinate the interaction. That structure lets incoming and outgoing audio continue without one direction unnecessarily blocking the other.
For developers, this is the important distinction: Gemini 3.8 Live is not just a model that turns speech into text and then reads an answer aloud. Google is presenting it as a foundation for agents that can maintain a conversation while invoking tools, handling visual context and responding to events in real time.
Where Google lists access to the models
Google lists Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking through Google AI Studio and the Gemini API. The company also lists the Gemini Enterprise Agent Platform as a developer and enterprise access surface.
Additional access depends on the Google product. Google lists Search Live and selected Workspace experiences for the models, and lists Gemini App access for Extended Thinking. Google Workspace Live capabilities also cover Gmail, Docs and Keep for the consumer rollout, while business availability was described as coming soon without a firm date.
For US developers, Google AI Studio and the Gemini API are the clearest entry points named for both models. The model listing alone does not turn every Gemini conversation into a Gemini 3.8 Live session; access is tied to the product surface and model option being used.
What Google’s benchmark numbers measure
Google reports separate results for Gemini 3.8 Live Extended Thinking rather than one universal performance score:
- 82.6 on Artificial Analysis’ Speech to Speech Quality Index.
- 68.6% on τ-Voice.
- 35.1% on Sierra’s τ-Voice-banking benchmark.
- 97.7% on Big Bench Audio.
Google also reports that standard Gemini 3.8 Live placed second in the Speech Agent Arena. These figures come from different evaluations with different goals, so they describe separate aspects of voice-agent performance rather than a single ranking that combines all four percentages.
The practical takeaway is visible in the model split itself. Gemini 3.8 Live is the option Google presents for speed, scale and fluid dialogue. Gemini 3.8 Live Extended Thinking is the option for workflows where deeper reasoning and spoken progress matter more than keeping the interaction as lightweight as possible.
