AI Credits or Your Own Key? The Math Nobody Shows You

You open a tool's pricing page and find the line: "500 AI credits per month." Nobody explains what a credit is worth, how many credits a summary uses, or what happens on the 18th, when the balance runs out. The math is opaque by design.

There's another way to charge for AI, and it has an ugly acronym for a name: BYOK, "bring your own key." Instead of buying credits from the tool, you connect the key from your account with a model provider — OpenAI, Gemini, or Claude — and pay for usage directly to whoever processes it. This article compares the two models with a calculator in hand.

The two models, without the marketing

Credits: you buy units from the tool

The tool buys from the AI provider wholesale, packages the usage as "credits," "messages," or "AI actions," and resells it. The per-unit price includes its margin, and the unit almost never corresponds to anything verifiable: a short question and a summary of thirty comments can cost the same credit, even though the real processing cost is very different.

When the balance runs out, the feature stops. You buy an extra pack, wait for the month to roll over, or move up a plan. Leftover credits rarely carry over.

Your own key: you pay the provider

You create an account with the provider, generate a key, and paste it into the tool. From then on, each AI operation is charged per token processed, at the provider's list price, and appears on their invoice. The tool isn't part of that equation.

A token is the chunk of text the model reads or writes, roughly three-quarters of a word. Providers charge per million tokens, and the difference between models is large: lightweight ones cost a fraction of the price of the heavier reasoning models.

Criterion	Tool credits	Your own key (BYOK)
Who sets the price	The tool, with margin built in	The provider, on the public price list
When it runs out	The feature stops until you buy more	It doesn't: you pay for what you used
Model choice	Set by the tool	Yours, even by type of task
Spending visibility	Credit balance	Usage by day and by model, in the provider's dashboard
Spending limit	The pack you bought	A monthly cap you configure
Initial effort	None	Create an account with the provider and paste the key

Why credits cost more than they seem

Four effects add up.

The margin. It's legitimate — the tool takes on usage risk — but it exists and you don't see it. Every credit carries the provider's cost plus the tool's protection against heavy users.

The rounding. Charging per "action" ignores text size. Anyone doing many small operations — summarizing a comment, improving a title — pays as if each one were large.

The balance that expires. A monthly pack is paid in advance. In the month the team traveled, did little, and used half, the other half evaporated.

The hidden model. If the tool decides which model to use, it has an incentive to pick the cheapest when margins get tight. You notice it through a drop in quality, without knowing the cause.

When credits are the better option

Your own key doesn't always win. Three cases where the pack pays off:

• You don't want to set anything up. Creating a provider account, adding an international card, and generating a key takes fifteen minutes, but it's an extra step. Someone who'll use AI twice a month shouldn't spend that time.
• Your company can't open an account with a foreign provider. Some financial operations have strict vendor rules. In that case, credit built into the tool's invoice solves a real administrative problem.
• Volume is very low and the pack is already included. If the plan you were going to buy anyway comes with enough credits, there's no savings to chase.

How to estimate your usage before deciding

You can do the math in ten minutes, with no complicated spreadsheet.

1. List the week's operations. How many documents become tasks, how many discussion summaries, how many questions to the assistant. Be realistic: most teams do between twenty and a hundred operations a week.
2. Estimate the size of each one. A task summary with thirty comments sends a few thousand tokens. A follow-up question sends far fewer. A long document turned into a plan is the most expensive operation of the lot.
3. Pick the model by type of task. Summaries, titles, and checklists work well on lightweight models. Save the expensive model for what requires reasoning, such as breaking a large scope into tasks with dependencies.

With operation counts and the model defined, the provider's own calculator closes the math. The result is often surprising: for a small team with moderate daily use, monthly usage comes in well below the price of an equivalent credit pack — and drops even further when simple tasks run on a lightweight model.

What to check before choosing

1. Does the tool let you choose the provider, or does it only work with one?
2. Can you use different models for different functions, or is the choice a single one for everything?
3. Is the key stored encrypted?
4. Is there a log of what the AI did, with author and time?
5. Can you turn off AI without losing anything in the rest of the product?

The last two questions are the ones that most separate serious products from demos. AI that takes actions needs auditing and confirmation for anything irreversible, regardless of who pays the token bill.

The decision, in one sentence

If your team will really use AI, every day, your own key pays back the initial effort in the first month and gives you back control of the cost. If use is occasional, built-in credit is convenient and the difference doesn't matter.

What you can't do is choose without knowing which model the tool uses. Before signing up, ask. And if the answer is vague, treat that as part of the answer. If you want to go deeper on the subject, also read what AI really solves in a project and see where it fits into day-to-day work.