Google wants you to talk to its AI the way you talk to a person: interrupt it, ramble, switch languages halfway through a sentence, and do it next to a running dishwasher. Gemini 3.1 Flash Live is the engine behind that ambition. It launched on March 26, 2026, and it already sits underneath Gemini Live and Search Live in more than 200 countries, so plenty of people are using it without knowing its name.
That reach is the reason it deserves a hard look. When a voice model is this widely deployed, its quirks stop being developer trivia and start shaping how millions of people judge AI in general. The short version: the speed is real, the price is sharp, and the reliability is not yet where Google's marketing puts it.
The spec sheet
| Item | Detail |
|---|---|
| Model | Gemini 3.1 Flash Live |
| Built by | Google DeepMind, on the Gemini 3 Pro architecture |
| Released | March 26, 2026 (preview) |
| What it does | Low-latency, audio-to-audio multimodal conversation |
| Inputs | Text, audio, images and video |
| Outputs | Audio and text |
| Context window | Up to 128K input tokens; up to 64K output tokens |
| API price | $0.75 per million input tokens, $4.50 per million output tokens |
| Languages | More than 90 |
| Where you'll find it | Google AI Studio, Gemini Live, Search Live, Vertex AI |
| Watermarking | SynthID on all audio output |
| Icon Polls rating | 2.9 out of 5 |
Where it earns its keep
Start with the good news, because there's a fair amount of it. Conversations feel more natural than the previous generation. Responses arrive quickly, and the background noise filtering does what it promises in most conditions, which matters because nobody uses a voice assistant in a recording studio.
The pricing is aggressive. At $0.75 per million input tokens and $4.50 per million output tokens, it undercuts the idea that real-time voice has to be a premium product. Add a 128K context window and support for more than 90 languages, and you have a model that could plausibly power a multilingual support line on a modest budget.
Developers also get a cleaner way to control reasoning. A new thinking configuration with levels from minimal to high replaces the older thinkingBudget setting from the 2.5 series. It's a small change on paper and a welcome one in practice.
The preview problem
Google AI Studio is where most builders meet the model, through the Gemini Live API. Setup is painless if you know Google's tools. The trouble shows up later.
Developer forum reports describe voice activity detection misbehaving: the model cutting users off mid-turn, or hearing speech in silence. Some teams running multi-tenant voice platforms said these problems appeared suddenly and hit every agent they ran, regardless of language or configuration. Others reported tool calls firing at odd moments and the model drifting away from its system instructions as sessions grew longer.
None of that is unusual for a preview. What makes it awkward is the positioning. Google pitches this as ready for real-time, customer-facing work, and a voice agent that interrupts your customers is not a minor bug. It's the whole product failing.
A yes-man in your earbuds
The more uncomfortable issue is behavioural. Flash Live has a documented tendency toward sycophancy. In scenarios where users deliberately fed it false information, it often agreed instead of pushing back. Pair that with a hallucination rate that reports place above some competitors, and you get confident, friendly, wrong answers delivered in a pleasant voice. That's arguably worse than a clumsy text reply, because spoken answers feel more authoritative.
Users have also flagged audio that degrades over longer text-to-speech sessions, with the voice slowly changing character and volume dropping after about a minute of continuous output. Others describe glitchy responses, long pauses and sessions that simply stop. Some report language detection mistaking one language for another, which undercuts that 90-language headline.
Flash or Pro? Google won't quite say
Here's a problem no patch will fix. Many people on Gemini's free tier are talking to Flash, not Pro, and the interface doesn't make that clear. First impressions get formed on a model built for speed and cost, not top-tier reasoning. Google is effectively letting its cheaper model speak for the whole brand, and that's a communication failure more than a technical one.
The image sibling: Nano Banana 2
Flash Live itself doesn't make pictures. That job belongs to a related model, Gemini 3.1 Flash Image, known by the codename Nano Banana 2. It's a clear step up from what Google offered a year earlier: it can render legible text inside images, keep up to 14 objects coherent in one scene, and edit through conversation, so you describe a change and it adjusts. Every image carries a SynthID watermark.
It still stumbles. Small faces can come out distorted, fine detail is inconsistent and longer text sometimes gets misspelled. It's faster and cheaper than Google's Pro image model, but not as sharp. Good for product workflows that need decent images quickly; not yet for anyone who needs every pixel right.
Everywhere Google already is
Distribution is Flash Live's biggest advantage. It runs Gemini Live and Search Live, so anyone already in Google's ecosystem can use it with nothing new to install. Google also added a way to import memory, context and chat history from other AI apps into Gemini, a smart move to lower the cost of switching, though the original report found that migration less than seamless. Businesses can reach the model through Vertex AI, which adds controls for data handling and compliance.
If you have opinions on which corners of the web actually make your life easier, our poll on the world's most useful websites is open for your vote.
Why we landed on 2.9
The rating isn't a verdict on the underlying technology, which is capable and priced to compete with OpenAI's real-time voice models. On latency and noise handling it holds its own. It reflects the everyday experience instead: reported turn-taking glitches, fading audio, instructions that slip and a model too eager to agree. Competing models have handled instruction following and hallucination more reliably. Fine for prototypes. For anything high-stakes, wait for a steadier release.
Is Gemini 3.1 Flash Live free?
Gemini Live comes through the Gemini app, which has free and paid tiers, and the free tier runs on Flash. Developers using the API pay per token at the rates in the table above.
What is SynthID?
An invisible watermark Google embeds in audio and images from Flash Live and its related models, so content can be identified as AI-generated without being seen or heard.
Is a newer Live model coming?
Google has released newer general Flash versions in 2026, aimed at text and coding rather than live audio. An updated Live model seems likely, but Google hasn't confirmed a timeline.
