Reading an AI bill.

A while back I ran one simple test that changed how I think about AI costs.

Same bill, different angle: Tokens are a budget explains where the multiplication starts inside a workflow.

One task — short prompt in, a few hundred words out. I priced the identical job across models at ten thousand runs. The cheapest option came to about a dollar. The most expensive came to over two hundred.

Same task. Same shaped answer. A 200× difference, decided entirely by a dropdown.

Here’s the whole pricing model, in operator terms. You pay for what goes in and what comes out, measured in tokens — pieces of text, roughly three-quarters of a word each. Input tokens are everything you send: the prompt, the pasted email, the case history, the instructions. Output tokens are what it writes back. Each model has its own price per million of each, and the differences between models aren’t percentages — they’re multiples.

Which means an AI bill is just three numbers multiplied: how much text per task × how many tasks × which model. That’s it. That’s the whole bill.

And each number is a lever you control. Send less and better-chosen context. Question the run count — does this need to fire on every ticket, or on the ones a cheap filter flags? And match the model to the task, because a £0.0001 job done by a £0.02 model is a 200× overspend that looks like a rounding error until it runs a million times.

The uncomfortable operator truth: the model dropdown is a procurement decision, and in most teams nobody is treating it as one. The engineer picks whatever was used in the demo. Finance sees one aggregated line called “AI”. The 200× hides in between.

Five minutes with your top workflow’s three numbers will tell you more than a quarter of vendor meetings.

Try this

Try it: calculate the cost of a workflow, then use Model Switch to compare candidate models on the tasks that matter.