The conversation need not stop for the task
Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September. The first is aimed at responsive, cost-conscious voice applications. The second is designed for more demanding reasoning and multi-step work.
The practical change is parallel activity. Google says its models can keep a conversation going while tools run in the background. Extended Thinking can also speak while reasoning, offering an acknowledgement or progress update instead of leaving the user in silence.
That could make a spoken assistant less awkward to use. It also creates a small but important design problem: a reassuring voice may be heard before the work is done. An acknowledgement, a plan and a successful action are three different states.
The application still has a job to do
Google's Live API documentation describes a continuous connection that carries audio and other inputs, rather than a succession of separate text prompts. Users can interrupt the model. Tools can connect the conversation to outside information and programs.
The tool guide makes a distinction that is easy to miss in a demo. Developers define the available functions, and their application handles the tool responses. The voice model does not remove the need for software that executes an action, catches a failure and returns the result.
Consider a booking assistant as a hypothetical example. Saying that it is checking availability should not be confused with making a reservation. A useful interface would distinguish the search, any request for permission and the confirmed booking. These are implementation questions, not claims about a product Model Current has tested.
Access is rolling out, not uniform
Google says both models began rolling out through the Gemini API and AI Studio. Its enterprise offering starts in private preview. The announcement places 3.8 Live in Search Live and Extended Thinking in Gemini Live, with subscription-dependent access in Workspace.
That is a rollout announcement, not confirmation that every account, country and application already has identical access. The product-specific entitlement should be checked before a team builds a workflow around it.
The accompanying model card says both variants accept audio, images, video and text, and produce audio and text. It describes them as based on Gemini 3 Pro. Those are model specifications, not a guarantee that each consumer interface exposes every capability.
Fluency is not a reliability test
Google reports strong results on voice and task-completion evaluations. Model Current has not independently reproduced those results or tested the models. Benchmark performance should not be treated as a measured success rate for a particular customer service or workplace process.
The company's own model card lists hallucinations, occasional slowness and timeouts. It also gives a January 2025 knowledge cutoff. A live conversation does not, by itself, make the underlying knowledge current.
For a pilot, the useful tests are ordinary ones: interrupt a request, let a tool fail, change the instruction halfway through, and check whether the assistant reports what actually happened. Faster conversation is a real usability goal. Clear confirmation still needs to be built around it.
Sources
- Google — Gemini 3.8 Live launch15 September 2026 announcement: two models, parallel speech/tool work, claimed evaluations and staged product access. Company claims, not Model Current test results.
- Google DeepMind — Gemini 3.8 Audio model cardPublished 15 September 2026. Model lineage, supported inputs and outputs, hallucinations, latency/timeouts and January 2025 knowledge cutoff.
- Google AI for Developers — Live API overviewContinuous streaming, interruption and application integration. The general language-count figure differs from the launch announcement and is not used.
- Google AI for Developers — Live API tool useFunction declarations and application-managed tool responses. Older-model support table is not used to infer new-model capabilities.
