Google's Gemini 3.8 Live Adds Extended Thinking for Voice Agents That Reason While Speaking

Google announced Gemini 3.8 Live and a companion model, Gemini 3.8 Live Extended Thinking, on September 15, 2026. According to iFynx, writing on dev.to, the release targets production voice agents and represents a move away from cascaded pipelines that convert speech to text, reason over it, and then synthesize a reply, toward native speech-to-speech systems.

Google's developer materials, as summarized by iFynx, position the new models for agents that need asynchronous function calling, near-real-time visual grounding, precise handling of alphanumeric strings such as confirmation codes, and mid-conversation language switching across more than 97 languages. 9to5Google, cited in the post, reports the models are rolling out across Gemini Live, Gmail Live, Docs Live, Keep Live and Search Live.

Two models, two latency profiles

Rather than a single voice model, Google shipped a deliberate pair. Gemini 3.8 Live is tuned for fluid turn-taking, cost efficiency and visual grounding, and is described as built for scale. Gemini 3.8 Live Extended Thinking is tuned for multi-step reasoning that continues while the model narrates its progress out loud. Google claims the Extended Thinking variant ranks first on Artificial Analysis' speech-to-speech leaderboard, though iFynx flags this as a vendor-reported figure that should be treated as indicative rather than settled.

The developer documentation adds a configuration difference: Extended Thinking exposes a thinking_level setting with low, medium and high options, while the standard Live model uses an interleaved profile without that control. iFynx's reading is that the standard model suits direct tasks where responses must feel instant, while Extended Thinking fits flows where the agent has to plan, wait on slow tools, or reason over complex state.

Speech continues while tools execute

The capability iFynx highlights most is asynchronous function calling: tools run in the background while audio keeps streaming. Extended Thinking leans on early spoken cues and live progress narration so that pauses never read as failure to the user.

This shifts client-side engineering, iFynx argues. Google's Live API guidance tells clients to track interactionStatus, which reports IN_PROGRESS or IDLE, rather than relying on turnComplete alone, because intermediate speech can arrive with turnComplete set to true while reasoning is still underway. A client that switches its UI to a listening state too early will interrupt the agent mid-thought. Function declarations in Extended Thinking sessions should also specify "behavior": "NON_BLOCKING", since blocking tools conflict with how the architecture streams.

Grounding, pricing and language coverage

The Live models can process live visual input in near real time, which iFynx sees as useful for camera-assisted support, document verification and in-app help. Google's listed audio pricing for the Live API is $0.005 per minute for input and $0.018 per minute for output. iFynx considers that affordable enough to make broad voice features budgetable, but recommends tracking cost per completed task rather than per session, since long Extended Thinking calls with continuous narration compound quickly.

On languages, Google claims mid-conversation switching across 97-plus languages. iFynx, which builds Arabic-first and bilingual products, treats automatic detection as helpful but unfinished, and recommends explicit confirmation of critical data such as amounts and account numbers, plus separate measurement of code-switching failures rather than relying on averaged language scores. A related Gemini 3.5 Transcribe release, covering 85-plus languages with a reported streaming word error rate around 4.0 percent, is cast as a transcription layer for analytics and captions, not a substitute for speech-to-speech agent design.

Access runs through AI Studio and partners including LiveKit, Pipecat, LangChain, Vercel and Agora. iFynx reads that packaging as an honest admission that the model is not the whole system: teams still own WebRTC behavior, barge-in policy, offline recovery and consent flows for microphone and camera access.

Why it matters

The release does not invent voice agents, iFynx notes, but it pushes continuous, tool-using conversation toward becoming the default expectation. Two consequences stand out. First, model selection becomes a per-workflow design decision: a fast dialogue model for direct interactions, a reasoning model for high-value multi-step ones, with unit economics measured against outcomes. Second, the plumbing around the model, from status tracking to non-blocking tool declarations to scripted progress speech, now determines whether an agent feels collaborative or broken. Teams that choreograph speech, tool execution and status signaling as a single experience are the ones most likely to ship voice agents users trust.