Google launches Gemini 3.8 Live voice models that reason while talking

Google released two new voice models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, on 15 September 2026, built for real-time speech-to-speech voice agents. Google said the models are aimed at production-grade use cases such as customer service and assistants.

Gemini 3.8 Live Extended Thinking scored 82.6 on Artificial Analysis' Speech to Speech Index, placing first ahead of OpenAI's GPT-Live-1 (Astra, medium) at 81.5 and xAI's Grok Voice Think Fast 2.0 High at 81.3, according to MarkTechPost.

The Extended Thinking model scored 68.6% on the Tau-Voice benchmark, up from 30.1% for the standard 3.8 Live model, and 97.7% on Big Bench Audio, Google said.

Time to first audio is 1.18 seconds for 3.8 Live and 1.35 seconds for Extended Thinking at its high reasoning setting, compared with 2.99 seconds for Gemini 3.1 Flash Live, according to the notes.

Rather than pausing to think silently, the Extended Thinking model acknowledges a request out loud, keeps talking, runs tools in the background and reports results once finished, TechRepublic reported. The model offers low, medium and high reasoning levels.

Both models are generally available in the Gemini API and Google AI Studio as hosted services only. Audio input is priced at $0.005 a minute and output at $0.018 a minute, and the models can switch between 97 languages mid-call. All audio output carries a SynthID watermark.

Extended Thinking is also coming to Gemini Live, Docs, Gmail and Keep for subscribers, Google said. Agora, LangChain, LiveKit, Pipecat and Vercel support the new models.

Google did not give a date for when Extended Thinking will reach Gemini Live, Docs, Gmail and Keep beyond describing it as coming to those products.