Google's New Gemini 3.8 Live Lets Voice AI Think While Talking

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, bringing new native speech-to-speech AI models to the Gemini API and Google AI Studio. The release combines multilingual conversations, live visual context and background tool execution, with Extended Thinking adding configurable reasoning for complex requests.

For developers, the important distinction is between keeping a conversation responsive and completing work that requires deeper analysis. Google now offers separate models for those demands. Choosing between them requires looking beyond the headline benchmark, particularly when calculating the cost of a long-running voice session.

Two Models for Different Conversation Demands

Both models accept audio, text, images and video. A conversational application can therefore combine spoken requests with camera context, rather than requiring users to describe everything they see. Google lists a 131,072-token input limit for each model.

Their intended roles differ:

• Gemini 3.8 Live, accessed as gemini-3.8-live, is Google’s default option for low-latency voice agents. It supports interleaved reasoning and asynchronous function calling, but developers cannot configure its thinking level. Google positions it for direct requests and responsive dialogue.
• Gemini 3.8 Live Extended Thinking, accessed as gemini-3.8-live-extended-thinking, performs background reasoning while maintaining spoken interaction. Developers can choose low, medium or high reasoning levels; minimal is not supported. Its intended workloads include complex troubleshooting and coordinated, multi-step tasks.

The standard model should not be described as having no reasoning. The difference is how reasoning operates and how much control developers have over it. That distinction matters when an application needs quick exchanges most of the time but occasionally encounters a request requiring substantial investigation.

Background Tool Calls Change the Voice Workflow

Asynchronous function calling allows an agent to request external information without suspending the entire conversation. The application still executes the function and returns its result: the Live API requires developers to handle tool responses in their client code.

For example, a booking assistant could acknowledge a request while its application searches flight and hotel services. Extended Thinking can provide spoken progress updates during background processing, then deliver a substantive answer after the necessary information arrives. This is the workflow Google’s thinking documentation is designed to support.

That behavior creates an important integration change. With Extended Thinking, turnComplete: true means an utterance has finished, not necessarily that the overall task is complete. Clients must continue processing messages and use interaction_status to distinguish IN_PROGRESS from IDLE. Otherwise, an interface could mistakenly stop listening for the result after an intermediate update.

Extended Thinking also requires non-blocking tool declarations. Setting behavior: BLOCKING produces an error, whereas standard Gemini 3.8 Live retains blocking execution for compatibility.

Google additionally highlights incremental content updates, allowing structured application data to accompany real-time audio. This is useful when the conversation needs newly retrieved information rather than only the context supplied when the session began.

97 Languages, With Live Translation Kept Separate

Google says Gemini 3.8 Live automatically detects and switches between 97 supported languages during a conversation. The capabilities documentation also allows developers to constrain the response language through system instructions.

That could simplify multilingual support workflows, but language coverage should not be mistaken for evidence of identical quality across every language or accent. Google also claims improved handling of alphanumeric information such as confirmation codes and claim numbers. Those are precisely the details developers should include in language-specific acceptance tests.

Live translation is a separate product choice. Google’s dedicated gemini-3.5-live-translate-preview model supports more than 70 languages and translates continuously as someone speaks. Unlike a conversational agent, it does not support tools or system instructions.

The distinction is practical: use a conversational model when the system should understand requests and act; evaluate the dedicated translation model when it should function as an interpreter. A multilingual assistant and a continuous translation pipeline are not interchangeable interfaces.

The Benchmark Lead Has Specific Boundaries

Extended Thinking’s state-of-the-art claim has independent support, but it applies to a particular evaluation. Artificial Analysis’ launch-period results place its High configuration first on the composite Speech-to-Speech Index, ahead of the following competitors.

The index equally weights speech reasoning, agentic performance, human preference and task success. Models need results for all four components to qualify. Consequently, the leading score does not establish that Extended Thinking