Gemini 3.8 Live Speaks While Reasoning, Complicating Async State

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, with developer access through the Gemini API and Google AI Studio and staged distribution across its enterprise and consumer products. Updated on September 17, Google’s launch post describes the Extended Thinking variant reasoning, making asynchronous function calls and delivering spoken acknowledgments or progress updates before its work is finished.

The September 15 release therefore changes what completion means for applications using Extended Thinking. The model is available to developers under a stable identifier, but a client can no longer interpret the end of one spoken response as proof that background reasoning, tool work and later audio have ended.

Extended Thinking leads the launch benchmarks

The two models target different workloads. Gemini 3.8 Live emphasizes low-latency dialogue and cost efficiency, while Extended Thinking adds configurable low, medium and high reasoning levels for planning and multi-tool tasks. Its ability to acknowledge a request before completing that work is the basis for the title’s claim that it can reason while speaking.

In the published results, Gemini 3.8 Live Extended Thinking scored 82.6 on the Artificial Analysis Speech to Speech Quality Index and completed 68.6% of tasks on the agentic τ-Voice benchmark. DataCamp’s independent launch analysis corroborates those scores and the September 15 release date, while noting that the speech-quality lead shown over the nearest competing systems was narrow.

The measurements answer different questions. The Speech to Speech Quality Index combines aspects of voice-agent quality, whereas τ-Voice measures completion of multi-step tasks involving tools. Neither score establishes how the model will perform with a particular language, function set, latency profile or failure mode, so the figures are comparative launch evidence rather than a guarantee for every production deployment.

One request can now contain several spoken turns

Extended Thinking can emit a short acknowledgment, continue reasoning, request tools and then produce more audio within the same interaction. A receive loop that closes after the first apparent answer may consequently miss a function call, an intermediate update or the final result.

The key distinction is between an utterance and the complete interaction. With Extended Thinking, turnComplete: true ends the current utterance; it does not necessarily mean the server is idle. An incoming interaction_status of IN_PROGRESS means processing, background reasoning or asynchronous tool work can continue, while IDLE means the overall task has finished.

This is why asynchronous state gets harder even though the conversation can sound smoother. Playback must remain ready for later audio frames, the tool dispatcher must preserve the active request, and the interface should not return to an idle or listening state until the interaction—not merely one segment of speech—is complete.

The migration contract is more than a model-name change

Google’s Extended Thinking model reference documents the stable identifier gemini-3.8-live-extended-thinking, a 131,072-token input limit and a 65,536-token output limit. It also specifies that functions must use asynchronous non-blocking execution, that blocking mode returns an error, and that sending new client content with turn_complete=true immediately interrupts active generation.

An integration adopting Extended Thinking therefore needs four state-management changes:

Keep receiving server messages after turnComplete, including later audio and function calls.
Use interaction_status as the lifecycle signal and return to idle only after IDLE.
Declare every function with behavior: NON_BLOCKING and preserve the association between asynchronous calls, their responses and the active user request.
Handle new client content as a possible interruption of current generation rather than assuming it will simply queue another independent turn.

Those requirements apply specifically to Extended Thinking. The regular Gemini 3.8 Live model retains the simpler behavior in which turnComplete: true closes the turn and returns the session to idle, so choosing between the variants also chooses a lifecycle contract.

The model’s feature boundaries remain firm

Extended Thinking accepts text, images, audio and video and can produce text and audio. It supports the Live API, audio generation, thinking, search grounding and asynchronous function calling, but it does not support caching, code execution, file search, structured outputs, URL context, image generation, Google Maps grounding or the Batch API.

That exclusion list matters for migrations built around more than voice. An existing pipeline that depends on structured output, cached context or code execution cannot preserve its architecture by changing only the model identifier; those functions must remain elsewhere in the application or use another supported model.

Availability also varies by channel. Developer access began through the Gemini API and Google AI Studio, while Gemini Enterprise started in private preview and other enterprise placements were announced as forthcoming. Consumer distribution began across Gemini Live and selected Workspace experiences with subscription-dependent access, rather than as one universal release.

The release trades simpler orchestration for continuous speech

The benchmark results make Extended Thinking a plausible option for voice agents that must plan, wait on tools or coordinate several actions. Its defining benefit is continuity: the user can hear an acknowledgment or progress update instead of waiting through every reasoning and tool-execution step in silence.

The cost of that continuity sits in the client protocol. The model may finish speaking without finishing the request, and it may speak again after a tool response arrives. The release is available to developers, but enterprise and consumer distribution remains partly staged; the next meaningful evidence will be whether its narrow benchmark leads and more complex interaction lifecycle hold up across production tools, languages, latency and recovery conditions.