One model for volume, one that thinks while it talks
Google released two speech-to-speech models on September 15, 2026. Gemini 3.8 Live is built for scale and cost; Gemini 3.8 Live Extended Thinking is built for harder tasks and reasons and speaks at the same time, using spoken cues such as "Let me check that" and narrating progress while tool calls run in the background. Both process visual input in near real time and switch among 97 languages mid-conversation. [1]
Developers got both through the Gemini API and Google AI Studio. Enterprises got a private preview in Gemini Enterprise, with its customer-experience product listed as coming soon. For consumers, the standard model powers Search Live, while Extended Thinking powers Gemini Live, live voice in Gmail and Keep for Google AI subscribers, and in Docs for AI Pro and Ultra subscribers. All generated audio carries a SynthID watermark. [1]
First place by a single point
Artificial Analysis's speech-to-speech leaderboard placed Gemini 3.8 Live Extended Thinking at high reasoning first, with an index score of 82.6, and 68.6 percent on its voice agent task test. OpenAI's GPT-Live-1, paired with GPT-6 Astra at medium effort and measured in a single trial, followed at 81.5. [3][1]
Atlas interpretation: A 1.1-point lead on a composite index, against a rival measured once, is a narrow first place. The more durable difference is architectural: Google puts the reasoning inside the voice model, while GPT-Live-1 hands it to a separate text model the developer picks. [1][3][4]
Sources
- Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
Google · Sep 15, 2026
- Gemini 3.8 Live Extended Thinking powers Gemini Live, Gmail, & Keep
9to5Google · Sep 15, 2026
- Speech to Speech Models and Providers Analysis
Artificial Analysis · Sep 22, 2026
- Build more natural voice experiences with GPT‑Live‑1 in the API
OpenAI · Sep 10, 2026