Gemini 3.8 Live and 3.8 Live Extended Thinking add real-time voice, camera vision, and live narration to Google's AI. Here's what's different.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are Google's newest real-time voice models, and the pitch is simple: talk to AI the way you'd talk to a person, interruptions and all. One is built for speed at scale, the other for multi-step tasks that don't break the conversation.
Google showed both off through Google AI's account on X, demoing Search Live, the camera-plus-voice feature, walking someone through a hands-on repair job in real time.
Voice assistants have promised "natural conversation" for close to a decade, and most still fall apart the moment you interrupt them or change topics mid-sentence. That's the actual problem Google is going after with this pair of models, not just another chat mode bolted onto the app.
Google split the launch into two models because "fast, natural voice chat" and "reason through a multi-step task" are different jobs. Cramming both into one model usually means neither works well.
| Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking | |
|---|---|---|
| Built for | Speed, scale, cost | Complex, multi-step reasoning |
| Interruptions | Mid-sentence | Mid-sentence |
| Languages | 97, switches on the fly | 97, switches on the fly |
| Visual input | Yes, camera-based | Yes, camera-based |
| Best for | Quick questions, live translation, Search Live | Planning, multi-step projects, "thinking out loud" tasks |
Point your phone's camera at whatever's broken, a bike chain that's slipped off, a pipe that's dripping under the sink, and describe the problem out loud. Search Live looks at the same thing your camera sees and talks you through the fix step by step, adjusting instructions as your hands move and the picture changes.
That's a genuinely different interaction than typing "how to fix a bike chain" into a search bar and reading a wall of text while your hands are covered in grease. Whether it holds up outside a demo video is the real question, and I'd want to try it on my own messy garage project before calling it solved.
The interesting part isn't that Extended Thinking is "smarter." It's that it talks while it works instead of going quiet. Most reasoning models make you wait through an awkward pause before answering. Extended Thinking is built to narrate: what it's checking, what it ruled out, while it works through something like an event plan behind the scenes.
If that holds up in daily use, it fixes the most annoying part of voice AI: the silence where you can't tell if it's thinking or if it just didn't hear you.
Honestly, I'd wait for hands-on testing before betting on this replacing how you talk to your phone day to day. Google's demo reel always looks smoother than the real product feels a week after launch. Ninety-seven languages sounds great on a slide, but models that actually handle accents, background noise, and someone talking over the assistant mid-answer are rare. That said, splitting a fast model from a "thinks out loud" model is the right call. It's the same lesson OpenAI and Anthropic have both landed on: one model can't be fast and deep at the same time.
For now, this looks worth trying if you're doing something hands-on and don't want to put your phone down mid-task: cooking, repairs, assembling furniture. For a plain back-and-forth chat, typing still gets you an answer faster.
Gemini 3.8 Live lands in a year where Google has moved fast on releases, from Gemini 3.6 Flash skipping straight past 3.5 Pro to Gemini 4 pre-training kicking off. For the wider picture of what else launched recently, our August 2026 AI model releases recap covers the rest.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Expo's new Exact framework runs one Contract app natively on web, macOS, iOS and Linux in a ~20KB runtime, built for AI coding agents to write, run and test it.
Run Qwen3.8-27B on your own GPU using llama.cpp's new v1.0.0 release and Unsloth's Dynamic v3.1 GGUFs, with exact commands and quant picks.
Researchers found 24,664 Lovable apps leaking passports, payments, and 200K+ chat messages via open Supabase databases. Here's the 10-minute RLS check.
Comments