Gemini 3.8 Live sees, speaks, and acts within a single session

Google launches Gemini 3.8 Live and Live Extended Thinking, two voice models capable of analyzing a camera feed, switching languages, and using tools without pausing the conversation.

A person points a phone at a leaking pipe. Gemini examines the area on camera, suggests an initial check, then adjusts its instructions based on the user’s answers and actions. The conversation can continue while the model consults an external service or prepares the next step.

Google is making this continuity the defining feature of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, introduced on September 15, 2026. The first prioritizes speed, scale, and cost. The second devotes more compute to complex operations while keeping the voice conversation going as they run.

These two models are not simply transcription systems followed by text-to-speech. They receive audio streams, text, images, or video and generate audio directly. Both are based on Gemini 3 Pro, with a context window of up to 128,000 tokens and a stated maximum output of 64,000 tokens.

Google presents this direct approach as an alternative to cascaded architectures. Those systems typically connect speech recognition, a text-based model, and a speech synthesis engine. Each handoff can add latency and lose information about pacing, intonation, or interruptions. Gemini handles the conversation within a single multimodal model, although the surrounding application remains responsible for audio transport, tools, and session management.

Gemini 3.8 Live is Google’s recommended option for most low-latency voice agents. It can be interrupted mid-sentence, detect when the user starts speaking again, and abandon the response it was generating. This prevents users from having to wait through an entire monologue before correcting a misunderstanding or clarifying a request.

The model can also switch languages within the same conversation. The Live API documentation lists 97 supported languages and says selection happens automatically. A user can begin a sentence in French, continue in English, and request an answer in a third language without opening a new session.

Support does not guarantee equal quality across all 97 languages. At launch, Google does not provide detailed results comparing comprehension, accents, interruptions, or response accuracy for each one. The list describes technical compatibility rather than a uniform level of performance.

Google also claims greater accent consistency and improved accuracy with alphanumeric sequences. This is intended for confirmation codes, claim numbers, technical references, and other information that can be difficult to distinguish in speech. The company does not publish a dedicated error rate for this capability, making it impossible to measure the improvement over previous generations.

Video adds what the camera sees to the conversation. Google shows Search Live guiding users through a bicycle chain repair or helping identify the source of a leak. Other demonstrations have Gemini play chess from a live view of the board or guide an employee through an unfamiliar interface.

The phrase “real-time video” requires some qualification. In the developer API, the visual stream is transmitted as a sequence of images, with a documented maximum of one image per second. The model can follow an evolving situation, but it does not receive every image captured by the camera as a conventional video system would.

That frequency may be sufficient to read a screen, identify a mechanical component, or notice that an object has moved. It is less suitable for very fast gestures, brief movements, or situations in which the exact order of several actions matters. A confident spoken instruction is also no guarantee that the model correctly interpreted what it saw.

Gemini 3.8 Live can call functions without blocking the conversation. An application can give it access to a calendar, document repository, booking system, or business tool. The model acknowledges the request, continues the exchange, and incorporates the result when the external service responds.

This asynchronous execution changes how a voice agent can be designed. A tool call has traditionally produced silence or a waiting message. In non-blocking mode, several operations can take place while the model asks a follow-up question or explains what is happening.

The Live API also accepts incremental context updates. An application can add structured data while the audio stream continues, such as the contents of a customer file, the status of a reservation, or a partial tool result. The model can incorporate that information into subsequent responses without rebuilding the entire conversation.

Continuity still depends on the complete infrastructure. Model latency alone does not determine the final experience. The user’s connection, media transport, speech detection, the external service, and the software playing the audio can each introduce delays or failures.

Gemini 3.8 Live Extended Thinking offers the same core functions but targets tasks that require more background reasoning. Its effort level can be set to `low`, `medium`, or `high`. A minimal setting is not available, while Gemini 3.8 Live does not let developers select a reasoning level manually.

Google describes Extended Thinking as reasoning and speaking in parallel. It may begin with a short acknowledgment such as “Let me check that,” then provide progress updates while it breaks down the problem, calls functions, and assembles the final result.

These updates should not be confused with complete access to the model’s internal reasoning. The API can provide thought summaries, and the model can narrate its progress, but these are user-facing generated explanations. They are not exhaustive, verifiable records of every computation that led to the answer.

Google’s selected examples include turning a sketch into a React interface, coordinating a multi-step booking, and producing a business plan with accompanying marketing materials. In the first case, the user comments on the result by voice while the model modifies the component. In the second, the conversation continues while several availability checks take