Gemini 3.8 Live for voice agents

Voice interfaces succeed when conversation continues naturally through interruptions, visual context and tool work. Google positions Gemini 3.8 Live as the default model for most low-latency agent experiences. Its September release describes near-real-time visual input, automatic language transitions across 97 supported languages, and tool or API calls that can execute in the background while a conversation continues. These capabilities invite use in troubleshooting, customer support and guided workflows, but a builder still has to supply tools, permissions and safeguards around real-world actions.

The model documentation makes a useful boundary clear. Gemini 3.8 Live accepts text, image, audio and video input and outputs text and audio, but it does not generate images, support structured outputs or use file search directly. Its model code is `gemini-3.8-live` and the listed input limit is 131,072 tokens. Google’s Live API guide notes that this model uses interleaved reasoning without a configurable `thinking_level` in session setup. A developer migrating from Gemini 3.1 Flash Live needs to change the model name and revisit asynchronous tool behavior and turn handling.

For tasks involving multi-step background analysis, Google offers a different endpoint, `gemini-3.8-live-extended-thinking`, that can continue speaking intermediate updates while it reasons and calls asynchronous tools. It should not be conflated with the standard low-latency model. Google reports strong benchmark and arena performance, but those numbers come from specific setups and cannot guarantee a particular live customer experience. API pricing has a free tier with different data-use terms and metered paid rates; the listed paid standard rates include $3 per million audio input tokens and $12 per million audio output tokens. Cost and latency should be measured on complete sessions, including video frames, tool calls and interruptions.