Google launched two new voice AI models on September 15: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The models target developers and enterprises building voice agents, according to a Google blog post.
The Big Picture
- Google released Gemini 3.8 Live for cost-efficient, large-scale voice agent deployment.
- Gemini 3.8 Live Extended Thinking targets complex, multi-step reasoning tasks spoken aloud.
- Both models automatically detect and switch between 97 supported languages mid-conversation.
- Gemini 3.8 Live Extended Thinking reasons and speaks simultaneously, narrating progress on background tasks.
- Partners including Salesforce, Genspark, and Lumeris are integrating the models into voice products.
Google Launches Gemini 3.8 Live
Gemini 3.8 Live processes visual inputs in near real-time, letting the model factor in what a user shows it while a conversation continues. The model also executes tool calls and API requests in the background without pausing dialogue, so it can acknowledge a request verbally while the task completes behind the scenes.
Gemini 3.8 Live Extended Thinking extends that design for harder problems. It reasons and speaks at the same time rather than pausing to think, using verbal cues such as “Let me check that” to bridge the gap while it works. For multi-step background tasks, it narrates progress aloud as each step finishes.
Google reports that Gemini 3.8 Live Extended Thinking scored well on independent benchmarks it cited, including a mark on Artificial Analysis’ Speech to Speech Quality Index, 68.6% on the τ-Voice agentic benchmark, 35.1% on Sierra’s τ-Voice-banking test, and 97.7% on Big Bench Audio. Google says Gemini 3.8 Live placed second in the Speech Agent Arena while remaining priced for high-volume use. On ServiceNow’s EVA-Bench, run through the Live API on the Gemini Enterprise Agent Platform, Google says its models balance accuracy against conversational quality for complex workflows.
— Google AI Studio (@GoogleAIStudio) September 15, 2026
Availability and Developer Access
Both models are rolling out starting today across several channels, per Google. Developers can access Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking through the Gemini API and Google AI Studio. Enterprises get Gemini 3.8 Live in private preview through Gemini Enterprise, with Gemini Enterprise for Customer Experience support coming later. Gemini 3.8 Live Extended Thinking follows the same enterprise path plus Google Workspace business accounts.
Consumer access varies by product. Gemini 3.8 Live reaches Search Live for general users. Gemini 3.8 Live Extended Thinking arrives in the Gemini Live app, for Google AI Pro and Ultra subscribers using Workspace tools in Docs, and for all Google AI subscribers in Gmail and Keep.
Voice platform providers including Agora, Fishjam, LiveKit, Pipecat, Vercel, and Vision Agents have integrated the Gemini Live API, handling the real-time media streaming layer so developers can focus on the voice interface itself rather than the infrastructure beneath it.
Every audio output from the new models carries a SynthID watermark, an imperceptible signal Google embeds directly into generated audio so the content stays detectable as AI-generated. Google points users to its model card for the full safety and responsibility approach.
Why It Matters?
The near real-time visual grounding and background task execution matter more than the benchmark scores for anyone actually building with these models. A voice agent that can glance at what a user is showing it, keep talking while an API call resolves, and switch languages mid-sentence solves the dead air and rigid turn-taking that made earlier voice assistants feel scripted rather than conversational. Extended Thinking’s approach, narrating progress instead of going silent to reason, is the more interesting design choice. It treats latency as a UX problem to manage rather than a technical constraint to hide.
Developers evaluating these models should treat Google’s cited benchmark figures as vendor reported until third-party leaderboards catch up, since independent scores for brand new releases often lag by days. Anyone already building on the Gemini Live API should check whether their integration handles the new background tool calling behavior correctly, since agents that assumed synchronous responses may need updates to work with a model that keeps talking while tasks run. Teams evaluating a switch should also confirm SynthID detection tooling covers audio in their moderation pipeline before relying on it for compliance purposes.
