Google Releases Gemini 3.8 Live for Voice Agents

Two voice models, with different ways to finish a task

Google's Gemini 3.8 Live launch announcement splits the rollout across products. Search Live gets the standard model, while Gemini Live gets Extended Thinking. Developers can access both through the Live API and AI Studio. Selected Workspace experiences are also receiving the reasoning variant, with access depending on the product and plan.

The distinction is more than a reasoning setting. Gemini 3.8 Live is Google's default for direct conversation where a quick spoken response matters. Extended Thinking is intended for requests that require several steps or slower tools. It can say that it is checking something, call a tool, and speak again while the overall request is still running. That makes a wait less silent, but a progress update is not the answer.

The session state is part of the product experience

Google's developer guide describes a concrete integration trap. With standard Gemini 3.8 Live, turnComplete: true means the model has finished its turn. With Extended Thinking, the same signal can mark the end of an intermediate utterance. The client must also watch interaction_status and treat IDLE as the point when the whole request is complete.

An app that ignores that difference could show a ready microphone or submit button while the assistant is still waiting for a tool result. It could also present a spoken status update as a final answer. Google requires non-blocking function declarations for Extended Thinking, so existing Live API applications need more than a model-name swap. The standard model has a simpler migration path from Gemini 3.1 Flash Live, although Google says its old thinking_level setup should be removed.

The new capabilities are real at the API level, but Google has not established that every app using Gemini Live will expose every tool or visual feature. A small independent browser demo by Simon Willison already shows that the models can support an interruptible voice conversation over Google's WebSocket API. It does not test Google's broader claims about task completion or compare the models under controlled conditions.

What a voice app sends and what it costs

These are hosted models. A device microphone can capture the audio, but the conversation reaches Google's API for inference. An app using live camera or screen context sends those selected inputs too. Calling a local app or service from the voice session creates another data boundary, so people handling sensitive material should check the application's microphone, camera, and connector permissions before relying on it. Our OpenAI GPT-Live privacy analysis explains the same data path: responsive audio does not imply on-device processing.

Google lists both new models under the same Live API price schedule. Its paid tier currently quotes $3 per million input audio tokens and $12 per million output audio tokens, with approximate per-minute figures of $0.005 and $0.018 respectively. The billing page says input audio can be charged while a proactive session is listening, and optional transcripts add text-token charges. A developer should budget for a real conversation length and transcription settings, rather than treating the headline audio estimate as a flat session price.

For a simple spoken command, the standard model avoids the extra lifecycle work. Extended Thinking fits a voice workflow that genuinely needs multi-step tool calls and can display progress without confusing it with completion. The next useful test is an end-to-end task in the actual app: start talking, interrupt it, wait for a slow tool, and check whether the interface stays accurate until the request is idle.