Google released two new live dialogue models on September 15, 2026: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. While incremental model numbers are common, the architectural shift in the Extended Thinking variant is not. It introduces parallel reasoning, allowing the model to process tasks in the background while streaming an audio response, fundamentally changing the interaction pattern for building complex voice agents.
what just shipped
Google has positioned the two models for different use cases. Gemini 3.8 Live is framed for scale, cost efficiency, and fluid dialogue. The more interesting model for developers building complex agents is Gemini 3.8 Live Extended Thinking. This version is designed for high-complexity tasks that require multi-step reasoning.
Both models are being rolled out across the Gemini API, Google AI Studio, and into products like Google Workspace and Search. The core capability upgrade is a move toward more natural and fluid voice interactions, but the mechanism for achieving this has direct implications for developers using the API.
asynchronous reasoning is the new default
The most significant change is how the Extended Thinking model handles work. It can process background tasks, like asynchronous tool calls, while simultaneously generating and streaming a continuous audio response. This is a departure from the traditional, blocking request-response cycle where the user waits in silence while the model completes its entire thought process.
For anyone who has built a voice agent, the benefit is obvious: the user gets immediate verbal feedback while the agent continues to work on a complex query. This makes the interaction feel less robotic.
However, it imposes a new burden on the client application. The developer documentation notes that when using asynchronous reasoning, a turnComplete: true signal no longer indicates that the model is idle. The server may still be processing tool calls or other background reasoning. Your client must continue to listen for subsequent server messages even after the initial audio stream for a given "turn" has finished.
updating your client-side logic
This new interaction model requires a shift in state management. A simple while loop that terminates on a completion flag is no longer sufficient. Your application needs to handle a more persistent connection and be prepared for out-of-band messages from the server.
Consider a conceptual Python client. Previously, you might have done something like this:
# Old pattern: simple request-response loop
response_stream = model.generate_content(audio_chunks, stream=True)
for chunk in response_stream:
play_audio(chunk.audio)
if chunk.turn_complete:
break # The turn is over, we can stop listening.
The new model requires logic that persists and handles different message types after a turn appears to be complete.
# New pattern: persistent listening for async events
response_stream = model.generate_content(audio_chunks, stream=True)
# Assume a persistent connection or long-lived stream
for message in response_stream:
if message.has_audio_chunk():
play_audio(message.audio_chunk)
if message.has_tool_call():
# Even if audio is playing or has finished for this turn,
# we receive and dispatch a tool call here.
dispatch_tool_call(message.tool_call)
if message.turn_complete:
# This no longer means the model is idle.
# It's just the end of this speech segment.
# The loop continues, listening for more messages like tool results.
print("INFO: Turn complete, but continuing to listen for background tasks.")
# The connection would remain open to receive further updates
This is a more complex, event-driven approach. The client has to be able to play audio while simultaneously listening for and processing other events, like function call requests from the model.
the so-what for builders
The introduction of asynchronous reasoning is a significant step toward more capable voice agents that can tackle complex, multi-step tasks without creating an awkward, silent user experience. It moves the interaction closer to a human-like collaboration where speaking and thinking can happen in parallel. For any team building production-ready voice agents, this is a pattern worth paying attention to.
However, this capability comes with the engineering overhead of more complex client-side logic. It also introduces an open question on cost. While competitors often publish per-minute pricing for their real-time voice APIs, Google has not yet done so for these new models, leaving a critical variable unknown for teams evaluating them for production use.
Ultimately, the release of Gemini 3.8 Live Extended Thinking signals that the frontier of voice AI is moving beyond simple, turn-based interactions. The next challenge for builders is to create applications that can gracefully manage the state of an agent that thinks and speaks at the same time.
Top comments (0)