Google releases Gemini 3.6 Flash and 3.5 Flash-Lite: What you need to know

Google just released two new Gemini models: 3.6 Flash and 3.5 Flash-Lite. There’s a third one too, called 3.5 Flash Cyber, but that’s only for security teams and government partners. So unless you work in cybersecurity, you probably won’t touch it directly.

These new models aren’t smarter in some flashy new way; they’re just more efficient.

AI models answer you by generating tokens (can be a single character, a punctuation mark, a part of a word, or even a whole word) one after another. Every token takes a bit of time and, if you’re using the model through an API, a bit of money too. So if a model can answer with fewer tokens, you get faster replies and a smaller spend. And that’s kind of the whole story behind this release.

Let’s go through what each model actually does, what the numbers mean, and who should care.

What is Gemini 3.6 Flash?

Gemini 3.6 Flash is Google’s new default model, replacing 3.5 Flash. It’s the one running behind most Gemini-powered apps and chatbots, even if you never see its name.

The main upgrade is that it just doesn’t need as many words to get to the right answer. Google says it uses 17% fewer output tokens on average than 3.5 Flash, based on tests from Artificial Analysis, an independent AI benchmarking group. On harder coding tasks, like the DeepSWE benchmark, the savings go up to 65%. Same job but a lot less text along the way.

It’s cheaper too: $1.50 per million input tokens and $7.50 per million output tokens, down from $9 per million output tokens on the older model. Put the lower price together with fewer tokens per task, and you get a real, noticeable drop in what it costs to run.

You might assume that using fewer tokens means worse answers. It doesn’t seem to work that way here. Google’s own benchmarks show gains across the board: coding accuracy went from 37% to 49% on DeepSWE, computer-use tasks went from 78.4% to 83% on OSWorld-Verified, and general knowledge tasks jumped from 1349 to 1421 points on GDPval. The model’s knowledge cutoff also moved forward, from January 2025 to March 2026, so it’s aware of more recent events out of the box.

What is Gemini 3.5 Flash-Lite?

If 3.6 Flash is the do-it-all model, 3.5 Flash-Lite is built for one thing: speed.

It can generate around 350 tokens per second, according to Artificial Analysis, which is fast even by Flash standards. It’s also the cheapest model Google has, at $0.30 per million input tokens and $2.50 per million output tokens.

This model isn’t designed for deep, complex questions. It’s meant for the kind of task that happens countless times an hour, like sorting support tickets, pulling data out of forms, or handling small steps inside a bigger AI agent.

Google says it beats the older 3.1 Flash-Lite by a wide margin, and in several tests, it actually outperforms the bigger 3 Flash model, scoring 54.2% versus 49.6% on SWE-Bench Pro, for example. Usually the smaller, cheaper model gives up some quality, but this time it’s beating models that cost more.

What about Gemini 3.5 Flash Cyber?

Flash Cyber is a version of 3.5 Flash trained specifically to find security bugs in code. It powers Google’s CodeMender tool, where several Flash Cyber agents work together to scan code, flag vulnerabilities, and suggest fixes.

You won’t be able to sign up for this one, though. Google is keeping it limited to governments and select security partners for now, and that’s on purpose: a model this good at finding security flaws could just as easily be used to exploit them.

How do the three models compare?

Why is Google focused on efficiency instead of just going bigger?

For the last couple of years, the go-to move for improving AI was to make it bigger. But bigger models cost more to run, take longer to answer, and start to strain under pressure when too many people use them at once.

So Google is trying something else with Flash. Instead of a bigger model, it trained one that solves problems in fewer steps, with less wasted text along the way. That’s cheaper for Google to run, and cheaper for anyone building on top of Gemini too.

What this means if your business runs on Google Cloud

If your company already uses Gemini through the API or Google Cloud, these two releases give you a real decision to make. You can send the simple, repetitive stuff to Flash-Lite, keep the harder reasoning on 3.6 Flash, and stop paying for power you don’t actually need.

Easier said than done, though. Figuring out which task belongs on which model, and setting the billing up so you actually see the savings, takes some know-how. That’s the kind of thing Revolgy, a Google Cloud Premier Partner, helps companies work through, i.e. matching the right Gemini model to the job and keeping cloud costs from quietly creeping up.

What’s next: Gemini 3.5 Pro and Gemini 4

Google also confirmed that a bigger model, Gemini 3.5 Pro, is already being tested with partners. It’ll roll out to everyone once testing completes, though there’s no exact date yet.

And in case you were wondering whether Google gave up on going big: work has already started on Gemini 4, its next major model. That one is still a way off, but it’s good to know it’s already in motion.

For now, Gemini 3.6 Flash and 3.5 Flash-Lite are already live in the Gemini app, Google AI Studio, Android Studio, and Google Antigravity. Flash-Lite is also rolling out in Google Search.