Google DeepMind has published Gemini 3.8 Audio, covering two new real-time models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are built for live multimodal interaction, accepting audio, images, video and text while returning audio and text.
The distinction is mostly about how much reasoning happens behind the conversation. Gemini 3.8 Live is positioned for lower-latency, high-volume voice interactions. Live Extended Thinking is designed for more complex, multi-step work where the model may need to reason or call tools while the audio session continues.
Google's model card says both are based on Gemini 3 Pro. It lists an input context of up to 128K tokens and up to 64K tokens of output for the audio models.
Extended Thinking changes how a live session behaves
Google's developer documentation makes the architectural difference clearer. Standard Gemini 3.8 Live is aimed at direct voice interactions and fast tools. Extended Thinking can continue background reasoning and asynchronous tool calls while streaming audio responses.
That changes what developers have to track. With the Extended Thinking model, a signal that an utterance has finished does not necessarily mean the whole interaction is idle. Google tells developers to watch the session's interaction status because reasoning or tool work can continue after a spoken turn appears complete.
The model can also use conversational fillers while work continues in the background. The intended effect is to avoid long silent gaps when an agent has to search, calculate or call external tools. That is a product-design choice as much as a model capability: a voice system has to manage latency in a way a text interface does not.
For developers, the documented model code is gemini-3.8-live-extended-thinking. Google lists text, image, audio and video as supported inputs, with text and audio output. Function calling is supported asynchronously. Search grounding is supported; code execution, file search and image generation are not listed as supported capabilities for this endpoint.
The model card also puts limits around the headline
DeepMind describes Gemini 3.8 Audio as optimized for real-time dialogue and latency-sensitive tasks, but its performance and safety claims remain Google's own evaluations. The model card exists partly to spell out known limitations, mitigation work and intended use, and it notes that those documents can change as models are revised.
That matters because "live" can sound broader than the actual interface contract. These models are not simply a general Gemini endpoint with every tool switched on. Developers have to work within the Live API's supported data types, tool behavior and session lifecycle.
The release is therefore less about one headline benchmark than a clearer split in Google's voice-agent stack. Gemini 3.8 Live is the fast conversational path. Live Extended Thinking trades a more complicated session model for background reasoning and multi-step work.
For users, Google says the technology is intended to make voice interactions across Gemini experiences more fluid. For developers, the practical change is more concrete: there are now separate live models for immediate dialogue and deeper agent-style reasoning, and they need to be integrated differently.
Reporting notes