The honest math of AI subscriptions

Twenty dollars a month has become the default price of a serious AI subscription, and almost nobody explains what it buys. Under every chat app is the same raw material: models priced per million tokens, roughly words in and words out. A subscription is a bet the company makes on your average usage. Light users subsidise heavy ones, the price covers the spread, and usage limits exist to keep the bet survivable.

The flagship problem

A single-model company has a constraint it cannot escape: its flagship is its brand, so your messages go to expensive models even when they do not need one. The result is that the easiest messages, which are most of anyone's day, are served at close to the highest cost. That cost has to come back out of the subscription price. You are not paying twenty dollars because your questions are hard. You are paying it because everything is billed as if it might be.

What routing changes

Most messages are genuinely well served by models that cost a tenth as much or less, and the difference is invisible precisely because those messages are easy. A router sends each message to the cheapest model that will truly handle it and saves the frontier models for the moments that reward them. That single change rebuilds the cost base of the whole product, which is why Heuricity can include every frontier family and still start at five dollars.

It also changes what honesty looks like. Heuricity measures usage in what things actually cost, over a rolling window with a weekly ceiling, instead of opaque message counts. Messages routed by Auto count less against your allowance than hand-picking a heavyweight for everything, because they cost less. Anything you run on your own models or accounts counts nothing at all. And the account page shows the running total of what you saved.

The catch, stated plainly

There is a catch and it is better said than discovered. Windows exist. If you hand-pick the most expensive model for every message all day, you will meet one, and the app will tell you so and offer real outs: continue on your own connections, top up at cost plus a plain margin, or wait for the window to roll. What we will not do is silently swap you to a worse model and hope you do not notice. If a limit is doing its job, you deserve to see it working.