Google's new voice models can talk and reason in real-time
What's the story
Google has launched its latest speech models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, to tackle the latency issue of voice-based artificial intelligence agents. These are the company's most advanced voice processing models yet, capable of near-real-time reasoning and concurrent speech-and-thought processing. They can also perform third-party software tool calls in the background, as per Google.
Performance metrics
Top-tier benchmark results
The new models have achieved top-tier benchmark results.
The Gemini 3.8 Live Extended Thinking model scored a record 82.6 on the Artificial Analysis Speech to Speech Quality Index, beating out GPT-Live-1-Astra and Grok Voice Think Fast 2.0.
Meanwhile, Gemini 3.8 Live ranked second on the alternative Speech Agent Arena benchmark and topped ServiceNow Inc.'s EVA-Bench rankings with its impressive performance metrics.
Enhanced capabilities
AI agents can now multitask
The new models are designed to perform tool and application programming interface calls in the background, while keeping a natural pace of conversation with human users.
This means AI agents can continue chatting with their users while working on tasks assigned to them at the same time.
Google said this offers a more natural, human-like calling experience without constant interruptions as agents browse the internet for answers.
Model advancements
Models can understand and generate speech in 97 languages
The new models also come with advanced features such as automatic language detection and mid-conversation language switching.
They can understand and generate speech in 97 languages, support near-real-time visual grounding, and use early verbal cues like "let me check that" to acknowledge user prompts more naturally.
Google said these capabilities make conversations with AI agents feel more lifelike and human-like.
Model access
Gemini 3.8 Live available now via Gemini API
Gemini 3.8 Live is available now via the Gemini API and Google AI Studio, as well as an enterprise private preview in Gemini Enterprise and Search Live.
As for Gemini 3.8 Live Extended Thinking, it can be accessed through the same channels and also in Google Workspace via Docs, Gmail, Keep (for subscribers), and in the Gemini Live applications.
Developers can integrate these models within their apps through partner platforms like Vercel, Agora, LiveKit, Pipecat, Fishjam, and Vision Agents.
Misinformation detection
SynthID watermarking and costs
When audio files are generated by the new models, they will carry an invisible SynthID watermark. This can be used to detect misinformation, Google said.
The standard Gemini 3.8 Live model costs $0.005 per minute for audio inputs and $0.018 per minute for outputs.
The Extended Thinking model also charges for reasoning tokens and additional inputs like video and documents as part of its pricing structure.