Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
Google ● Covered by 6 sources
Google’s new Gemini voice models are out, with one tuned for scale and one for harder tasks. They can talk while they work, which is the whole point of a voice agent.
Based on reporting by Google — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind has launched two new Gemini voice models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are built to make real-time conversations with AI feel more natural, but they split the job in different ways. One is aimed at scale and cost efficiency. The other is aimed at heavier reasoning and more complex work.
The company is pitching 3.8 Live Extended Thinking as the more capable option for enterprise tasks. Google says it took the top spot on Artificial Analysis’ Speech to Speech Quality Index with a score of 82.6, and led agentic task completion on τ-Voice with 68.6% and on Sierra’s τ-Voice-banking benchmark with 35.1%. It also scored 97.7% on Big Bench Audio, while still staying at a competitive price point versus other frontier models.
3.8 Live, meanwhile, is the more practical workhorse. Google says it ranked second in the Speech Agent Arena and is designed to be highly cost-effective for developers and enterprises building voice products. On ServiceNow’s EVA-Bench, the company says both models pushed the Pareto Frontier for complex workflows by balancing accuracy with conversational quality.
The technical pitch is pretty clear: these models don’t just answer, they keep the conversation moving while work happens in the background. Gemini 3.8 Live can process visual inputs in near real time, switch automatically among 97 supported languages mid-conversation, and handle tool and API calls without stopping the chat. Extended Thinking goes a step further, speaking and reasoning at the same time, with lines like “Let me check that...” and live progress narration for multi-step tasks.
Google is rolling them out across a lot of its own products too. 3.8 Live is starting in the Gemini API and Google AI Studio, with private preview for enterprises in Gemini Enterprise and availability in Search Live. 3.8 Live Extended Thinking is also starting today in the API and AI Studio, with private preview for enterprises, plus access in Gemini Live and in Workspace for certain subscribers. Google says all audio it generates is watermarked with SynthID so it can be detected later.
My take — AI-written commentary, not fact-checked reporting
This is the right direction for voice AI: less chatbot theater, more doing the thing while talking about it. The industry has spent years pretending “talking naturally” was the hard part, when the real test is whether the model can keep its mouth shut long enough to finish the task in the background. Google is finally leaning into that, which is refreshingly unsexy.
Read more about this at: Google