Gemini 3.8 Live grabs the voice-agent crown from OpenAI and xAI

Google has pushed ahead of OpenAI and xAI in the race for real-time voice agents, releasing two models on 15 September that can reason and run tools mid-conversation instead of going silent to think. The move lands squarely in the market for customer-service bots and voice assistants, where latency and the ability to actually do something while talking have become the competitive line.

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are native speech-to-speech models built for production use. The standard version is tuned for speed and cost at scale; Extended Thinking adds multi-step reasoning at low, medium and high settings, with the model acknowledging a request out loud, continuing the conversation, running tools in the background, and reporting back once done.

The numbers back the positioning. On the Artificial Analysis Speech to Speech Index, Extended Thinking (High) scores 82.6, 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. On Tau-Voice, Extended Thinking hits 68.6% against 30.1% for the standard model, and Big Bench Audio comes in at 97.7%. Time to first audio is 1.18 seconds for Live and 1.35 seconds for Extended Thinking High, against 2.99 seconds for the prior Gemini 3.1 Flash Live.

Pricing is $0.005 a minute for audio in and $0.018 a minute for audio out, with mid-call switching across 97 languages. The models are generally available, hosted-only, through the Gemini API and Google AI Studio, and are launching with support from Agora, LangChain, LiveKit, Pipecat and Vercel. Extended Thinking is also coming to Gemini Live, Docs, Gmail and Keep for subscribers. All output carries a SynthID watermark.

The benchmark lead is narrow but real, and it arrives at a moment when voice has become the next battleground after text and code. By closing the gap between "thinking" and "talking," Google removes the dead air that has made voice agents feel unnatural against a live human agent, a functional edge that pricing alone wouldn't deliver.

The infrastructure partners matter as much as the scores: Agora, LangChain, LiveKit, Pipecat and Vercel give Google immediate distribution into the stacks where voice-agent builders already work, putting pressure on OpenAI and xAI to match both the latency and the tooling ecosystem, not just the benchmark. Enterprises building customer-service or assistant products now have a lower-latency, competitively priced option with hooks into the frameworks they already use, a combination that could pull integration decisions toward Google before rivals respond, as TechRepublic notes.