Google's new Gemini Flash models are built for AI that runs itself
Google DeepMind's 3.6 Flash and 3.5 Flash-Lite promise cheaper, faster AI agents. Here is what that means for the apps and games you actually use.
Key points
- Google DeepMind released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, aimed at running AI agents at scale.
- 3.6 Flash uses 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, priced at $1.50 per million input tokens and $7.50 per million output tokens.
- Gemini 3.5 Flash-Lite runs at 350 output tokens per second, priced at $0.30 per million input tokens and $2.50 per million output tokens.
- On OSWorld-Verified, which tests an AI's ability to use a real computer, 3.6 Flash scored 83.0% against 78.4% for 3.5 Flash.
- Google DeepMind confirmed pre-training has started on Gemini 4, its next flagship model.
Google just quietly changed the maths for anyone building an AI that does things for you.
On release day, Google DeepMind rolled out two new versions of its Gemini family: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Gemini is Google's line of AI models, the software brains behind chatbots and, increasingly, AI agents (programs that carry out multi-step tasks on your behalf). The headline isn't a bigger, smarter monster model. It is the opposite. These are the workhorses.
Why should a normal person care about a cheaper AI model?
Price and speed decide whether the clever demo you saw last year actually shows up inside the apps you use.
Running a large AI model costs money every time it answers. That cost is measured in tokens, chunks of text the model reads and writes. Gemini 3.6 Flash uses 17% fewer output tokens than 3.5 Flash to do the same job, according to the independent Artificial Analysis Index. On one coding test called DeepSWE, the saving hit 65%.
The model rambles less and gets to the point faster. That drops the bill for the company building your app, which is how features like an in-game AI companion or a smart customer service bot become financially possible. Pricing sits at $1.50 per million input tokens and $7.50 per million output tokens.
The little sibling might be the real story
Gemini 3.5 Flash-Lite is the small, fast one. Google clocks it at 350 output tokens per second, priced at $0.30 per million input tokens and $2.50 per million output tokens.
That speed matters for anything that needs to feel instant: an AI opponent reacting in real time, or a document-processing tool chewing through thousands of invoices overnight.
On Terminal-Bench 2.1, a test of how well an AI can drive a computer command line, Flash-Lite jumped from 31% to 54%. OSWorld-Verified, which asks the AI to operate a desktop as a human would, came in at 74.0%. Both new models now include "computer use" as a built-in tool, meaning the AI can click buttons and fill in forms without a developer wiring it up by hand.
When we covered Flash Cyber's debut on 21 July, the story found the model punching well above its weight on vulnerability hunting. That context matters here: Flash-Lite follows the same philosophy of doing more with less.
What about safety?
Google says 3.6 Flash ships with stronger safeguards against misuse in chemical, biological, radiological and cyber-offence scenarios. It has also been trained to refuse fewer harmless requests, a common complaint with earlier Gemini versions that sometimes said no to everyday questions.
A specialist cousin is also in the mix. Flash Cyber powers CodeMender, Google's agent that hunts for security holes in software code.
What happens next?
Gemini 3.5 Pro, the bigger flagship model, is still testing with partners. Pre-training has begun on Gemini 4.
The honest read on this release: efficiency gains this size are rarely glamorous, but they're the thing that actually gets AI features shipped to real users at scale. Frontier models get the headlines. Flash models get the work done.
