The short version
Gemini 3.8 Live and 3.8 Live Extended Thinking are Google's "most advanced live dialogue models yet," designed to make voice interactions more natural, fluid, and intelligent[1]. The headline upgrade is that these models can reason and speak at the same time — they don't make you wait in silence while they think. For developers, they're the building blocks for production voice agents; for everyone else, they power the Gemini app, Google Workspace, and Search[1].
Two models, two jobs
Google split the release into two tiers that serve different needs[1]:
- Gemini 3.8 Live — built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding. It's the workhorse for high-volume voice agents.
- Gemini 3.8 Live Extended Thinking — built for high-complexity tasks, with increased intelligence and multi-step reasoning. It reasons and speaks simultaneously, narrating progress while it works through background steps.
If you're building a customer-facing voice bot that needs to be cheap and responsive, start with 3.8 Live. If you need an assistant that can walk through a multi-step booking or build a functional component from a sketch while talking to you, that's Extended Thinking territory[1].
Benchmarks: top of the speech charts
Google's claims for these models are backed by published third-party and internal benchmarks[1]:
| Benchmark | Result | Model |
|---|---|---|
| Artificial Analysis — Speech-to-Speech Quality Index | 82.6 (#1) | 3.8 Live Extended Thinking |
| τ-Voice (agentic task completion) | 68.6% | 3.8 Live Extended Thinking |
| Sierra τ-Voice-banking | 35.1% | 3.8 Live Extended Thinking |
| Big Bench Audio (reasoning) | 97.7% | 3.8 Live Extended Thinking |
| Speech Agent Arena | #2 | 3.8 Live |
Results as reported in Google's Gemini 3.8 Live announcement, September 2026[1].
On ServiceNow's EVA-Bench, Google says the two models "push the Pareto Frontier" for complex voice-agent workflows by balancing accuracy with conversational quality[1]. The practical takeaway: Extended Thinking is a genuine step up for tasks that mix reasoning with spoken interaction, not just a marketing label.
What it can do
- 97 languages, switched on the fly. The models automatically detect and transition between 97 supported languages mid-conversation[1].
- Real-time visual context. 3.8 Live processes visual inputs in near real time, so you can point a camera at something and talk about it[1].
- Background tool use. It executes tools and API calls in the background while continuing the conversation — acknowledge the request, keep chatting, and let tasks finish[1].
- Transparent reasoning. Extended Thinking uses early verbal cues ("Let me check that…") and live progress narration so you're not staring at a silent model[1].
- SynthID watermarking. All generated audio is watermarked with SynthID to keep AI output detectable and reduce misuse[1][4].
Where to get it
Both models are rolling out now, with different channels per tier[1]:
- Developers: the Gemini API (Live API) and Google AI Studio.
- Enterprises: private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience coming soon.
- Everyone: 3.8 Live in Search Live; 3.8 Live Extended Thinking in Gemini Live, and in Google Workspace (Docs for Google AI Pro and Ultra subscribers; Gmail and Keep for all Google AI subscribers).
For consumers, the simplest way to try it is Gemini Live in the Gemini app, or through a Google AI subscription[1]. Developers can find integration details in the Live API docs[3].
Pricing
Google positions 3.8 Live as "highly cost-effective" for scale, and says Extended Thinking maintains "a highly competitive price point compared to other frontier models"[1]. Billing runs through the Gemini API on Google's standard tier structure — free tier for experimentation, then paid and enterprise tiers[2]. Specific per-token and per-second rates are listed on the official Gemini API pricing page, which you should check before committing to a production build[2].
Bottom line
Verdict: If you're a developer or enterprise building voice agents, Gemini 3.8 Live is worth a serious look — the benchmark numbers and the background-tool design solve real problems (latency, silence-while-thinking, language switching). For everyday users, it's a free upgrade inside the Gemini apps you may already pay for. There's no urgent reason to switch from a text-first workflow unless you genuinely want to talk to your AI.
Frequently asked questions
What's the difference between Gemini 3.8 Live and 3.8 Live Extended Thinking?
3.8 Live is tuned for speed, cost, and fluid conversation; Extended Thinking is tuned for high-complexity tasks with multi-step reasoning, and it speaks while it thinks[1].
Can Gemini 3.8 Live understand images?
Yes — it processes visual inputs in near real time, so it can ground a conversation in what the camera sees[1].
Is it available on my current Google plan?
Consumer access runs through Gemini Live, Search Live, and Google Workspace. Docs support requires Google AI Pro or Ultra; Gmail and Keep are available to all Google AI subscribers[1].
Sources
- Google — "Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking" (Sep 15, 2026) — blog.google
- Google AI for Developers — Gemini API pricing — ai.google.dev
- Google AI for Developers — Live API — ai.google.dev
- Google DeepMind — Gemini 3.8 Audio model card — deepmind.google