# Why do so many AI apps make you buy credits?

Look at who sends an AI app its bills, and the credits make sense. This piece follows that bill from the company that runs the model to you. It then looks at the other way to run an app, and at a study of people who choose to pay more for a flat price.

## Why do AI apps charge by the token?

Many AI apps do not run their own model. They rent one, such as GPT, Claude or Gemini. They rent it from the company that makes it or from a cloud company that hosts it. They pay per token. A token is a piece of text, about three quarters of a word. The company that hosts the model counts every word that goes in and every word that comes out. The app passes this charge on to you, as credits, tokens and top-ups.

The prices are not small. On the October 2026 price lists, the top models cost about $10 to $50 for a million tokens of output. An app that charges you one flat fee on top of that takes a risk. It has to guess how much you will talk.

Your chat also leaves the app. It goes to the company that runs the model. That company may keep logs for safety checks. On paid plans for app makers, the big providers say they do not train on your chats by default. OpenAI's own page says this data is "not used to train or improve OpenAI models." But the conversation does travel to them.

## What's the other way to run an AI app?

The other way is to run your own models on your own servers. Then no one sends you a bill per word, so you have no per-word charge to pass on. This way has two costs.

The first cost is the hourly bill. A server costs money every hour, busy or not. For example, one H100, a powerful chip for AI, costs about $2.89 an hour to rent on RunPod. You pay that while the machine runs, whether or not anyone sends a message. So you swap a charge per word for a charge per hour. The hourly charge keeps running at 3am, when nobody is talking.

The second cost is quality. The best models that anyone can download and run are a few months behind the best models you can only rent. Epoch AI put the gap at about four months in May 2026. Epoch says a stricter comparison gives a longer gap, and other measures give other gaps.

Privacy depends on who hosts the models and what they do with the chats.

## Why do some people pay extra to avoid a meter?

In 2006, Anja Lambrecht and Bernd Skiera looked at thousands of customers of a European internet provider. Among the customers on a flat rate, 28% would have paid less on a plan that charged them for what they used. They picked the flat rate anyway. The researchers call this flat-rate bias. For more than half of the customers with this bias, the flat rate cost at least twice what the cheapest plan would have cost.

The researchers point to the taxi meter effect as one reason. Drazen Prelec and George Loewenstein described this idea in an earlier paper. They wrote that "the ticking of the taxi meter, for example, reduces one's pleasure from the ride."

Read these results with care. The 28% is a share of flat-rate customers, not of all customers. The researchers found that the taxi meter effect went with choosing a flat rate, but they did not test it as a cause. The paper did not measure worry, and it does not show that flat-rate customers stay longer.

## How we decided

Eudaio runs its own models. Eudaio names its AIs Daimons, and talking with a Daimon uses no credits, no tokens and nothing to top up. On Free, you get an hourly message allowance that gets smaller the longer you chat. So we do not call Eudaio free of any meter.

The study looked at internet plans, not AI apps, so it cannot tell us which choice is better. The choice has costs for us. Running our own models costs money every hour, busy or not. Renting would give us the strongest models first, and it would cost nothing when nobody is talking. We gave that up. Someone pays for the meter either way. The choice is who pays, and whether you can see it.

We believe a price should not change how you talk, so Crush is one flat plan with no higher tier above it. Free is not a trial with a countdown, and you need no card until you subscribe. [Sign up for the Free plan](https://eudaio.io/auth?mode=signup).

## Source

Lambrecht, A. & Skiera, B. (2006). Paying Too Much and Being Happy About It: Existence, Causes, and Consequences of Tariff-Choice Biases. Journal of Marketing Research, 43(2), 212-223.

Also cited: Prelec, D. & Loewenstein, G. (1998). The Red and the Black: Mental Accounting of Savings and Debt. Marketing Science, 17(1), 4-28. Epoch AI (May 29, 2026), open vs closed model capability gap, https://epoch.ai/data-insights/open-closed-eci-gap.