September 17, 2026
A voice assistant becomes more useful when you can refine a request while it is working. You might add a constraint, correct a detail or ask why it chose a particular option without having to restart the task.
On September 15, Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, expanding its voice AI lineup. The announcement brings a practical question into focus: how much work can an assistant accomplish while keeping a conversation moving? Google’s launch announcement.
Two models for different kinds of conversation
Google’s developer documentation distinguishes the models by the complexity of the work they are designed to handle.
| Model | Main emphasis | Example workload |
|---|---|---|
| Gemini 3.8 Live | Responsive dialogue and straightforward tasks | Voice search or a quick device command |
| Gemini 3.8 Live Extended Thinking | Background reasoning and longer sequences of actions | Investigating a technical issue using several information sources |
Extended Thinking can deliver brief spoken updates while processing a request. Its integration also distinguishes an individual spoken response ending from the overall task being complete. That distinction matters: an assistant can finish a sentence while still having work to do. Google’s guide to thinking in the Live API.
For users, the important question is whether this coordination makes a task easier to supervise. A reassuring voice is useful only if the assistant also makes its progress and remaining uncertainties clear.
What “reasoning during a conversation” means in practice
Imagine asking an assistant to arrange a meeting involving several people. A useful interaction might look like this:
- You describe who should attend and the preferred day.
- The assistant checks the relevant information.
- While it works, you add that one participant cannot attend before midday.
- It presents suitable options and asks you to confirm before sending invitations.
This is an illustrative workflow, not a claim that every Gemini account can perform it today. It would require access to calendars, appropriate integrations and permission to act.
The value would come from keeping the task understandable as your instructions evolve. Did the assistant incorporate the new constraint? Did it check everyone’s availability? Has it merely proposed a time, or actually booked it?
These questions become especially important when voice makes an interaction feel effortless. The application should still show the user what has happened.
The software around the model still does essential work
Google’s Live API supports function calling: the model can request an operation from software connected to it. The application handles that request and returns the result. Asynchronous tools allow an operation to run without blocking the conversation. Live API tool documentation.
For example, a voice assistant connected to a stock-management system could request an inventory lookup. The model would interpret the result and explain it; the inventory system would remain the source of the stock figure.
This is why access to a capable model does not automatically create a complete business assistant. A useful implementation needs dependable data, suitable tools and clear boundaries around what the assistant may change.
For a first deployment, a sensible scope would be a task with an easily checked outcome: locating a document, preparing a draft or retrieving an order status. Sending a message or modifying a customer record introduces a different level of responsibility.
Benchmarks reveal a trade-off worth understanding
Artificial Analysis reports the following results for the two models, as checked on September 17. Extended Thinking was evaluated at its High setting.
| Evaluation | Gemini 3.8 Live | Extended Thinking — High |
|---|---|---|
| Speech to Speech Index | 76.0 | 82.6 |
| Agentic performance — τ-Voice | 30.1% | 68.6% |
| Conversational dynamics | 96.1% | 91.9% |
The index combines several measures. τ-Voice tests completion of simulated customer-service tasks; conversational dynamics evaluates interaction behaviours such as pauses and interruptions. Artificial Analysis: results and methodology.
These results suggest a meaningful choice. Extended Thinking performs better on the task-completion benchmark, while the standard model scores higher on conversational dynamics. A single overall ranking would hide that difference.
Neither percentage predicts the success rate of your own application. A practical evaluation should include your users’ vocabulary, typical corrections, incomplete requests and tool failures. Measure whether the assistant reaches the correct outcome and how much human intervention it requires.
Where can you use the new models?
Google announced a rollout beginning September 15:
- Developers: both models through the Gemini API and Google AI Studio.
- Consumer products: Gemini 3.8 Live in Search Live; Extended Thinking in Gemini Live.
- Workspace: Extended Thinking in Docs for Google AI Pro and Ultra subscribers, and in Gmail and Keep for Google AI subscribers.
- Enterprise: private preview in Gemini Enterprise, with further business availability described as forthcoming.
These are the launch channels described by Google, not a guarantee that every account already has every feature. Google’s rollout details.
Before choosing a use case, check the actual product available to your account. Access through a developer API, a personal subscription and an employer-managed workspace can involve different setup and controls.
What the new models still cannot guarantee
Google DeepMind’s model card lists possible hallucinations, occasional slowness and timeouts. It also specifies a January 2025 knowledge cutoff. A recent launch therefore does not mean every answer comes from up-to-date internal knowledge. Gemini 3.8 Audio model card.
For information that changes frequently, the application needs an appropriate current source. A confident spoken answer should not substitute for checking a delivery date, a price or a policy against the relevant system.
There is also a design challenge: users need to distinguish an acknowledgement from a verified result. Hearing that the assistant is checking something should never be mistaken for confirmation that the action succeeded.
What this could change for everyday AI use
Our view is that the most useful test of conversational reasoning is how well it supports correction. Can someone change a requirement halfway through? Can they understand what remains unfinished? Can they stop an action before it has consequences?
If those interactions work reliably, voice could become a more practical way to coordinate work across software. Choosing the right model is one part of that process. The quality of the connected tools, the clarity of the interface and the user’s ability to verify the outcome will determine whether the experience saves time.
Sources
- Google — Introducing Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, September 15, 2026
- Google AI for Developers — Thinking in the Live API
- Google AI for Developers — Tool use with Live API
- Google DeepMind — Gemini 3.8 Audio model card
- Artificial Analysis — Speech to Speech models, results and methodology