AI-generated editorial illustration; not a documentary photograph.
Google has launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two speech models designed to reduce pauses while voice agents reason and work on tasks. Google says both models can process speech and thought concurrently, while running third-party tool calls in the background during a conversation.
What changes
The models are intended to let an agent continue speaking naturally while it browses or performs an assigned action, rather than stopping the conversation during each operation. Google also says they can detect languages automatically, switch languages mid-conversation and understand and generate speech in 97 languages.
Capabilities and claims
Google says Gemini 3.8 Live supports near-real-time visual grounding and can use early verbal acknowledgements before completing a request. The company reports that the Extended Thinking model scored 82.6 on the Artificial Analysis Speech to Speech Quality Index, while Gemini 3.8 Live ranked second on Speech Agent Arena and first on ServiceNow’s EVA-Bench. Google also reports Extended Thinking scores of 68.6% on T-Voice, 35.1% on T-Voice-banking and 97.7% on Big Bench Audio. These are company-reported benchmark results, not independent tests supplied by the source.
Availability and cost
Google says Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are available through the Gemini API and Google AI Studio. The models are also offered in an enterprise private preview through Gemini Enterprise and Search Live. Extended Thinking is additionally available to subscribers in Docs, Gmail, Keep and Gemini Live applications.
For the standard model, Google lists pricing of $0.005 per minute for audio input and $0.018 per minute for audio output. Extended Thinking also charges for reasoning tokens and additional inputs such as video and documents. Developers can access the models through partner platforms including Vercel, Agora, LiveKit, Pipecat, Fishjam and Vision Agents.
Audio provenance
Google says generated audio files include an invisible SynthID watermark intended to help detect misinformation. The source does not provide technical details about how the watermark is detected or its performance in real-world conditions.
Source: Google’s new speech model Gemini 3.8 Live supports real-time reasoning
SIGNAL TITANS
Leave a Reply