Google Launches Gemini 3.8 Live Voice AI Models

Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, introducing advanced real-time voice AI capabilities. The models are integrated into the Gemini app, Google Workspace, Search, the Gemini API, and Google AI Studio, with access provided to Google AI Pro and Ultra subscribers and private previews for Gemini Enterprise.

Gemini 3.8 Live is designed for scale, cost efficiency, and fluid dialogue with visual grounding. The Extended Thinking version focuses on high-complexity tasks and multi-step reasoning, allowing the AI to reason and speak simultaneously. This architecture enables the model to run tools in the background while continuing to speak, meaning the end of a spoken response no longer necessarily signals the completion of a task.

Both models support 97 languages and process text, image, audio, and video inputs in near real time. Google reports that Extended Thinking leads in agentic task completion on the τ-Voice and Sierra’s τ-Voice-banking benchmarks, and achieved a score of 82.6 on Artificial Analysis’ Speech to Speech Quality Index. To mitigate misinformation, Google is applying SynthID watermarking to all generated audio.

Audio pricing is set at $3 per million input tokens and $12 per million output tokens. While the models are stable, Google noted that the broader Live API remains in preview and may still experience timeouts, slowness, or hallucinations.