Google on Sept. 15, 2026, introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its latest models that take in speech and answer in speech, built for voice assistants that carry out tasks on their own. Both opened the same day to developers through the Gemini API, a connection that lets developers plug a service into their own apps, and in Google AI Studio, with rollouts starting in Search Live, Gemini Live and Workspace. Google is offering one model for smooth everyday conversation and another for harder, multi-step work, both built on the same Live system.
Gemini 3.8 Live is built for scale and low cost. It holds fluid spoken conversations, takes in what it is shown in near real time, switches among 97 languages mid-conversation, and can use other tools and services in the background while the conversation continues. Gemini 3.8 Live Extended Thinking is built for more complex tasks. Google says it can work through a problem while it is still speaking, using early spoken cues and live updates on its progress during multi-step work in the background.
The rollout puts Live into Search Live and Extended Thinking into Gemini Live, into Docs for Google AI Pro and Ultra subscribers, and into Gmail and Keep for all Google AI subscribers. Businesses get both models in private preview in Gemini Enterprise.
Google says Extended Thinking took the No. 1 spot on Artificial Analysis' Speech to Speech Quality Index with 82.6, a score the testing firm's own listing shows too, with standard Gemini 3.8 Live at 76.0. Google also says Extended Thinking led on tests of completing tasks by voice, with 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking, and scored 97.7% on Big Bench Audio, a reasoning test. On the firm's Speech Agent Arena, a ranking built from people's preferences in side-by-side comparisons, standard Live debuted with an Elo rating of 1083, behind only Gemini 3.1 Flash Live at 1096, while Extended Thinking ranked lower even as it led the quality index.
Developers call the models gemini-3.8-live and gemini-3.8-live-extended-thinking. Each can take in up to 131,072 tokens, the small chunks of text, often parts of words, that AI models read and write, and put out up to 65,536 tokens, and both support audio output. Google's developer post prices Gemini 3.8 Live audio at $0.005 a minute for input and $0.018 a minute for output, and Google says all audio its AI products make carries a SynthID watermark. It is not yet clear how far business access will widen beyond private preview, or how Extended Thinking's higher quality scores weigh against people's preference for the standard Live model.