Google launches Gemini 3.8 Live with simultaneous thinking and talking

The single most awkward part of talking to an AI voice assistant has always been the dead air. You ask it to check your schedule or look something up, and the line goes completely cold for five seconds while it spins in the background. You never quite know if it crashed or if it is actually working.

Google is rolling out an update aimed directly at fixing that robotic awkwardness. On September 15, the company launched Gemini 3.8 Live alongside a heavier-hitting variant, Gemini 3.8 Live Extended Thinking.

Simultaneous thinking, talking, and doing

Rather than forcing you to wait in silence while background tasks process, these new models can talk, reason, and run tools at the exact same time.

Gemini 3.8 Live handles visual inputs in near real-time, allowing you to point a camera at an object and get immediate commentary. It also automatically detects and swaps between 97 languages on the fly without requiring any manual toggling. When it needs to call an external API, it keeps chatting with you while the task finishes in the background.

The 3.8 Live Extended Thinking version takes this dynamic much further for complex enterprise workflows. Instead of freezing while calculating multi-step logic, it gives verbal cues like "Let me check that…" and narrates its progress as steps wrap up. It feels far less like querying an endpoint and much more like sitting next to a human assistant who keeps you posted while working.

The benchmark numbers back up the performance. The Extended Thinking model took the #1 spot on Artificial Analysis' Speech to Speech Quality Index with an 82.6 score. It also reached 68.6% on the τ-Voice benchmark, 35.1% on Sierra’s τ-Voice-banking benchmark, and scored 97.7% on Big Bench Audio. The standard 3.8 Live model secured second place in the Speech Agent Arena.

To address deepfake concerns, Google noted that all generated audio includes an imperceptible SynthID watermark directly in the audio output.

Both models are live today for developers inside the Gemini API and Google AI Studio, backed by integration platforms like LiveKit, Pipecat, Agora, and Vercel. For end users, the models are rolling out across Search Live, Gemini Live, and select Workspace apps like Docs, Gmail, and Keep.

Why it matters

We think this is Google tackling the single biggest bottleneck holding voice agents back from enterprise adoption: the trade-off between speed and deep reasoning.

Previously, voice models had to choose. They could be ultra-fast and surface-level, or smart and painfully slow. By letting the model narrate its thinking and run tools asynchronously without cutting off dialogue, Google solved the perceptual latency problem. It turns an annoying wait into an engaging conversation, which is critical for customer service, technical troubleshooting, and in-car assistants.

Just as important is the pricing pressure Google is putting on the rest of the industry. Independent tests and third-party breakdowns point out that while Gemini 3.8 Live performs right alongside rivals like GPT live and Grok voice, it does so at a fraction of the cost, running multiple times cheaper. Google is clearly weaponizing its infrastructure margins to lock developers into its Live API ecosystem before competitors can optimize their audio unit economics.