How FreeMilk Stays Free

FreeMilk is free. There's no subscription, no credits to buy and no per-token bill, and your daily allowance refills every day. AI is expensive to run, so it's fair to ask how. The answer is three decisions that keep our costs low enough for ads to cover them.

The short answer.

• Open models do the free work. FreeMilk's free mode runs on open-weight models such as GLM from Z.ai and DeepSeek. They cost a fraction of what the top closed models cost, and they've become remarkably capable.
• Ads pay for the tokens. Ads for developer tools, relevant to what you're building, fund your daily allowance.
• The agent runs on your phone. There's no cloud machine or sandbox to rent for your project. The agent works inside the FreeMilk app and writes your files to your iPhone.

Why AI apps usually cost money.

Two costs dominate. The first is tokens: every prompt, every file the agent reads and every line it writes runs through a model, and model makers charge by the token. The second is compute: many AI builders give each project its own machine in the cloud and pay for it while you work.

Both costs grow with every user and every prompt, which is why most AI tools sell credits or monthly plans. FreeMilk cuts the first and removes the second.

1. Our free mode runs on open models.

Open-weight models publish their weights, the trained parameters that make up the model. Anyone can download and serve them, so many companies compete to host them, and prices fall.

They're also good. Vals.ai's Vibe Code Bench has each model build a complete web app from a written spec, then runs automated tests on the result. DeepSeek V4.1-Flash, an open model, passed 85% of the tests and cost $0.41 per app. Claude Sonnet 5.5, the top-scoring closed model, passed 92% and cost $31.25 (Vals.ai).

Closed models still lead on the hardest problems (Vals.ai). When you want one, connect your ChatGPT subscription and use your plan in FreeMilk.

2. Ads pay for the tokens.

Your daily allowance is paid for by ads for high-quality developer tools. They're contextual: what you see relates to the project you're working on, so an ad is more likely to be a tool you'd actually use than a random banner.

Because open models cost so little, ads can cover the work. That's why there's no subscription and no per-token bill, and why your allowance refills every day.

3. The agent runs on your phone.

This is the most unusual part of FreeMilk. Most AI builders run your project on someone else's computer. FreeMilk's agent runs on your iPhone, inside the FreeMilk app, and writes your project's files to your phone.

That works because iPhone apps are already containerized. iOS keeps every app in its own sandbox and limits what it can read and write to its own container (Apple). The isolation that cloud builders rent a machine for, your iPhone already provides.

With no fleet of cloud machines to run, the biggest fixed cost of an AI builder disappears. Your files stay on your phone, every file is yours to read and change, and your work doesn't depend on someone else's server. Your prompts go to the model you picked so it can respond.

One step does leave your phone: compiling a native Swift app needs Apple's tools on a Mac, so FreeMilk builds those on remote Macs and streams the simulator back, as our iOS guide explains.

Free, by design.

We wrote about why this matters to us in our manifesto: the benefits of AI should be shared widely, and that takes abundant, free tokens. Open models, relevant ads and an agent that lives on your phone are how we keep FreeMilk free today.