Voice is quietly becoming the main battlefield in AI, and Google just fired the loudest shot of the season. On September 15, the company rolled out Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — a pair of models it calls its most advanced live dialogue models yet. If you have ever given up on a voice assistant because it froze the moment you mixed Hindi and English in one sentence, this launch is aimed squarely at you.
01.Two models, two very different jobs
The naming is confusing, so let me simplify. Gemini 3.8 Live is the workhorse — built for scale and cost efficiency, with fluid dialogue and visual grounding, meaning it can process what your camera sees in near real-time while you talk. Gemini 3.8 Live Extended Thinking is the heavy-duty sibling, designed for high-complexity tasks with multi-step reasoning. Think of the first as the friendly receptionist and the second as the senior analyst who happens to speak out loud while working.
02.It switches between 97 languages mid-conversation
This is the detail that made me sit up. The model automatically detects and transitions between 97 supported languages mid-conversation — no settings, no language picker. For a country where an average conversation can travel from Hindi to English to a regional language within one minute, that is not a gimmick, it is a feature built for us. The developer documentation also highlights accent consistency and accurate parsing of alphanumeric details like confirmation codes, which has always been where voice assistants fall apart.
03.It works while it talks
Here is the thing nobody is talking about enough: these models do not freeze while they execute tasks. Through what Google calls asynchronous function calling, 3.8 Live can run API and tool calls in the background while continuing to stream audio to you. Extended Thinking goes further — it uses natural verbal cues like "let me check that..." and narrates its progress live while multi-step background tasks finish. So you ask it to check something, it acknowledges, keeps chatting, and reads out the result when ready. That is a genuine shift from the awkward silence we have all tolerated since the first voice assistants.
04.The numbers, and what it costs
Google is not being shy about the benchmarks. Extended Thinking took the #1 overall spot on Artificial Analysis' Speech-to-Speech Quality Index with a score of 82.6, leads agentic task completion with 68.6% on the τ-Voice benchmark and 35.1% on the τ-Voice-banking benchmark, and scores 97.7% on Big Bench Audio. Both models carry a 128K token context window across audio, images, video and text. For developers, pricing is $0.005 per minute for audio input and $0.018 per minute for audio output — aggressive for what these models do. If you are building anything on paid AI APIs, metering that spend is a skill in itself; the cost optimization chapter in our GPT-6 Astra — Zero to Hero Guide covers the same budgeting discipline for the Gemini API.
05.Where you can try it today
The rollout is already live. Everyone can use 3.8 Live in Search Live on the Google app — tap the Live icon and just ask your question out loud, then keep asking follow-ups. It is also available in Gemini Live for Google AI Pro and Ultra subscribers, in Docs for Workspace business customers, and in Gmail and Keep for all Google AI subscribers. Developers get it in the Gemini API and Google AI Studio today, with a private preview running in Gemini Enterprise.
06.My honest take
Honestly, what surprised me most was not the benchmarks — it was how human the conversational flow sounds in the demos. The model acknowledging a request, running a task quietly, and narrating progress feels less like software and more like a colleague on a call. Whether that experience holds up on a patchy Indian mobile network is the real test. But between the 97-language switching, the visual grounding and the background task execution, this is the first time voice AI looks ready for the way India actually talks.
Been following this thread on the blog: Gemini Hacked Three Companies: Google's AI Just Had Its First Known Breakout and Gemini 3.8 Flash Cyber: Google's Most Powerful Security AI Isn't for Everyone.
Key Takeaways
- Gemini 3.8 Live and 3.8 Live Extended Thinking launched September 15 — Google's most advanced live dialogue models
- 97 languages, switched automatically mid-conversation, no configuration needed
- Asynchronous function calling means tasks run in the background while conversation continues
- #1 on Artificial Analysis' Speech-to-Speech Quality Index at 82.6; 97.7% on Big Bench Audio
- Free to try today in Search Live on the Google app
Frequently Asked Questions
What is Gemini 3.8 Live?
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are Google's most advanced live dialogue models, launched September 15. They switch between 97 languages mid-conversation and can run tool and API calls in the background while you keep talking.
How much does Gemini 3.8 Live cost developers?
Developers pay $0.005 per minute for audio input and $0.018 per minute for audio output. Anyone can try it for free in Search Live on the Google app.
How good is Gemini 3.8 Live Extended Thinking?
It took the #1 spot on Artificial Analysis' Speech-to-Speech Quality Index with a score of 82.6, and scored 97.7% on Big Bench Audio.
What This Means for You
If you are a student or an everyday user, open the Google app and try Search Live — it costs nothing and shows you where voice AI is heading. If you are a developer or a small agency in India, voice agents just became dramatically easier and cheaper to build, and the ₹150-crore call-center industry is going to be rebuilt around this technology within a couple of years. And if you are a professional wondering what to learn next, voice-first AI workflows — where you direct AI by talking instead of typing — are becoming a real skill. Our GPT-6 Astra — Zero to Hero Guide is a practical starting point for that exact skill set, from prompting fundamentals to building agents that do the work for you.