Gemini 3.8 Live Gets Extended Thinking. It Holds Background Context While You Talk

The new models no longer freeze when executing tools

Google announced Gemini 3.8 Live and version 3.8 Live Extended Thinking, which together form a new generation of conversational models. The key difference is the separation of fast responses from asynchronous tasks. While older models had to pause communication when working with tools and could not be interrupted, Extended Thinking runs in the background, allowing you to continue conversing with the model seamlessly even while it handles another task.

Although Google's primary page was somewhat limited in detail, the accompanying demos clearly illustrate the new architecture. The model combines a fast verbal stream (3.8 Live) and a separate backend process for structured reasoning (Extended Thinking) that do not interrupt each other.

For voice assistant UX, this solves the biggest bottleneck

Until now, voice agents worked great either as a chat or as task orchestrators, but never both at the same time. When you gave a complex prompt requiring data search or code execution, you had to stay quiet until the task finished. If you interrupted the model with a supplementary instruction, the entire chain-of-thought sequence often collapsed.

The division into two parallel streams changes this. It means developers can build agents that maintain small talk or accept new inputs while their backend part is still working on a heavy problem. This is a massive shift for any real-world assistant deployment on the go.

The architecture is more complex and demands perfect state synchronization

The limit lies in how well these two systems exchange context. If Extended Thinking in the background reaches a conclusion that is no longer valid because you verbally changed the instruction in the meantime, there is a risk of asynchronous hallucinations. The marketing claims the model handles multitasking smoothly, but real deployment outside of Google's demos will quickly reveal how aggressively the model discards old context branches upon new input.

At the same time, it is not yet clear how available Extended Thinking is via API for regular developers and at what cost, or if it is a proprietary feature locked solely in the Gemini app.

It will depend on whether parallel execution breaks under long API pauses

The proof that this works will not be another polished demo, but the ability to maintain a fluent conversation even when the background tool fails and has to make several retry loops. If the frontend model starts making things up just to fill the silence at that moment, the illusion of two parallel streams will rapidly fall apart.

Lilith's verdict

The AI assistant stopped being a walkie-talkie you have to wait on. Instead, it became a full-fledged colleague who can maintain normal eye contact while working on something complex.