Flat-Rate Model

A flat-rate model is a way of paying for a product. You transfer a fixed amount, usually per month, and are allowed to use the offering as often as you like in return. The price therefore doesn’t depend on how much you actually consume. This is familiar from mobile phone plans with unlimited data or from a streaming subscription for movies. The counterpart is usage-based billing: there, you pay per minute, per gigabyte, or per request. For AI services such as chatbots, the flat-rate model has since become the standard for private customers.

Why flat rates for AI services can become a loss-making business

With most digital products, an additional user costs the provider almost nothing. A song is recorded once and then played back millions of times without significant extra cost. With AI, it’s different. Every single response has to be computed by expensive specialized chips in a data center. This computing work costs electricity and hardware time, and it does so anew every single time.

This creates a problem that doesn’t exist in the same way with music or movies. A user who asks five questions a day is highly profitable for the provider. Another user who has code generated for hours on end can easily cause more costs than they pay per month. Sam Altman, the head of OpenAI, publicly admitted in 2025 that his then most expensive subscription, at 200 dollars a month, was running at a loss. The reason was precisely this effect.

Providers therefore calculate using averages. They hope that the many frugal users will help subsidize the few extreme heavy users. This calculation only works out as long as the customer base doesn’t shift. If word gets around among professionals about particularly cheap subscriptions, the mix can tip quickly.

Fair-use limits, throttling, and other emergency brakes

Almost no flat rate is truly unlimited. The terms of use usually contain a fair-use clause. It states, in effect: unlimited applies only to normal use. Anyone who consumes noticeably more than usual gets slowed down or blocked. With chatbots, you can see this in messages like “You’ve reached your limit, try again in three hours.”

Technically, providers work with several levers. A limit caps the number of requests within a time window. Throttling lets usage continue, but more slowly. And often, under heavy load, the system switches to a smaller, cheaper model that responds faster and requires less computing time. The user usually only notices this because the answers become shorter or less precise.

A hybrid form is also widespread. You pay a base fee for a fixed allowance, such as a certain number of image generations. Anyone who needs more buys additional packages. Experts call this a hybrid pricing model. It combines the predictability of the flat rate with the cost coverage of usage-based billing.

Flat rates in subscriptions, app stores, and corporate contracts

In everyday life, you encounter this model with practically every AI subscription. ChatGPT Plus, Google Gemini, or image generators typically cost 20 to 25 euros a month. In return, you get access to the better models and higher limits than in the free version. Coding assistants like GitHub Copilot work the same way.

For businesses and developers, the opposite is often true. Anyone who accesses an AI directly via a programming interface pays per processed text unit. These units are called tokens, and billing is done per thousands of them. This is risk-free for the provider, but hard for the customer to plan for. That’s precisely why the flat rate is so popular in the consumer business: you know at the start of the month what the service will cost.

In business news, the term usually comes up in connection with margins. Analysts then ask whether a provider’s subscription prices even cover the computing costs. A common misconception is the assumption that AI companies automatically profit from every subscription. With flat rates covering expensive computing power, that’s not true — prices are often set strategically to gain market share and are adjusted later.