What is usage-based billing? Tokens, credits, and outcomes, explained in Simplified Technical English - Arnon Shimoni

Skip to content

What is usage-based billing? Tokens, credits, and outcomes, explained in Simplified Technical English

Before I begin the ASD-STE100 part, let me explain:

I’m an airplane nerd (sometimes called "Avgeek"), and I really enjoy things that have to do with aviation. Well, aircraft maintenance manuals use a controlled language known as ASD-STE100. From here on out, I’ll try to use the rules to write about billing.

I will disclaim however, that I used Claude to help me with adherence to the rules after I had written the post.

ASD-STE100 has approximately 900 approved words. The sentences are short. Each sentence gives one idea. The voice is active. The standard exists because an unclear sentence in a manual can kill a person.

Pages about billing have the opposite problem. The words are large and the mechanism is hidden. The reader does not know who counts what, who pays for what, or who carries the risk.

This post obeys STE.

Words such as "token" and "credit" are technical names, and the standard permits technical names. All other text follows the rules. A controlled language can describe a hydraulic system. It can also describe a rate card.

What is usage-based billing?

Usage-based billing is a model where the price follows usage. When the customer does more, the customer pays more. When the customer does less, the customer pays less.

The system has 3 parts. A meter counts events. A rate card gives a price to each event. A ledger records the charges.

The parts stay the same in each version of the model.

Only one thing changes: the unit that the meter counts.

UnitThe meter countsWho carries the riskExamplesTokensModel input and outputThe customerAPI pricing at OpenAI, AnthropicCreditsActions, converted to an abstract unitSharedFigma, Lovable, CursorOutcomesThe result of the workThe vendorIntercom Fin, Chargeflow

This is not a standard diagram, but it is drawn to look like a schematic used in aviation maintenance.

What is a token?

A token is a small unit of text. A language model reads text as tokens and writes text as tokens. One token is approximately 4 characters of English. A short word is 1 token. A long word is 2 or 3 tokens. One page of text is approximately 500 tokens.

A tokenizer divides the text into tokens before the model starts its work. The model then writes the output one token at a time. The model provider counts the input tokens and the output tokens for each request. Output tokens usually have a higher price than input tokens.

Why are tokens important?

Tokens are the cost unit of AI. Each AI feature sends requests to a model. Each request creates a token count. The token count sets the cost of goods for the feature. Model providers set a price for each 1 million tokens. Different models have different prices.

Tokens are also the meter of AI. Electricity has the kilowatt-hour. AI has the token. A company that does not measure tokens for each customer does not know the margin for each customer. One AI company found that the top 5% of its users consumed 75% of its compute. The finance team found this at month-close, 3 months too late.

Some AI companies sell tokens to their customers at a markup. This is the simplest version of usage-based billing. The meter counts tokens. The rate card prices tokens. The invoice shows tokens.

Token pricing has a problem. The customer cannot predict the count. The customer does not know if a request costs 400 tokens or 40,000 tokens. The model decides. The bill arrives, and the customer learns the number after the fact.

Token pricing moves the cost risk to the person with the least control of it.

What is a credit?

A credit is an abstract unit that sits between the customer and the real cost. The customer buys a pool of credits before use. Each action in the product uses a set number of credits. One generated image costs 5 credits. One agent task costs 20 credits.

The vendor sets the exchange rate between credits and the cost. The customer sees the credit price. The customer does not see the token price behind it.

Credits are an architectural decision. The credit layer translates between two currencies: the cost to operate the product, and the price the customer agrees to pay.

Figma uses two types of credits. Some credits follow value. Some credits follow cost. The user decides if Gemini 3 Pro is worth 3 times the credits of a cheaper model. Figma moved that cost calculation to its customers. It is not my favourite design.

Why do credits give companies predictability?

Credits give predictability to 3 parties at the same time.

The customer gets a stable price for each action. The image costs 5 credits today. The image costs 5 credits next month. The customer can make a budget. The token count behind the image can move, and the customer does not care.

The vendor gets a stable margin structure. The vendor...