Two speech-to-speech models, including an "Extended Thinking" variant that takes first place in the Artificial Analysis index with 82.6, and automatically detects 97 languages.
Google launched two native voice models on September 15, 2026: Gemini 3.8 Live, cost-oriented and scale-based, and Gemini 3.8 Live Extended Thinking, designed for multi-step reasoning tasks. These are word-to-word models, without intermediate transcription, capable of executing tools in the background without interrupting the dialogue.
The novelty claimed on the Extended Thinking variant is to reason and speak simultaneously, instead of pause the time to reflect. She ranks first in the quality index of the artificial analysis with a score of 82.6, ahead of GPT-Live-1-Astra and Grok Voice Think Fast 2.0; the standard model ranks second to the Speech Agent Arena while remaining significantly cheaper. Both models automatically detect the language among 97 and process the visual in near real time to contextualize their responses.
The announced rate is US$ 0.005 per minute of audio input and US$ 0.018 per minute of output. Availability is immediate via the Gemini API and Google AI Studio for developers, with advance private access on Gemini Enterprise for companies, and on the general public side in Search Live as well as in Workspace, Gmail and Keep for Pro and Ultra subscribers. All generated audio bears an imperceptible SynthID marking.
Read the full article on the original source
Go to the source →