What does a million tokens actually cost

What is a token?

The unit in which a language model reads and writes text. Before a model does anything with your text, a program (the tokenizer) cuts it into pieces from a fixed vocabulary. Common words are one token, longer or rarer words are cut into parts, and punctuation and spaces count too. The providers give as a rule of thumb that a token in English is about four characters, roughly three quarters of a word. Dutch has longer words and less room in the vocabulary. So a Dutch text usually costs more tokens per word than the same text in English.

So the model does not count in words, characters or pages, but in tokens, and so does the bill. A page of text of five hundred words is roughly seven to eight hundred tokens; a long document of fifty pages roughly forty thousand. Code counts differently from prose, because symbols and indentation each cost tokens. Knowing exactly is only possible with the model's own tokenizer; most providers offer a counter for it.

Every provider has its own tokenizer, which means the same text costs more tokens with one model than with another. Two models with the same price per million are therefore not equally expensive for your text. That is the first of the price traps further on.

• A token is a piece of text from a fixed list; about four characters of English.
• Dutch usually costs more tokens per word than English; code counts differently again.
• Every provider cuts differently, so the same text does not cost the same everywhere.

Why do input and output have a different price?

Because they cost the model different work. Input (everything you send the model: your question, the instructions, the document, the earlier parts of the conversation) is processed by the model in one go, in parallel. Output (what the model writes back) is created token by token, and each next token requires a new calculation over everything that came before it. That makes output much more expensive for the provider to produce, and you see that in the price. With most models, a million output tokens cost a multiple of a million input tokens.

For Claude Fable 5.1 the ratio on 5 September 2026 is one to five: $10.00 input against $50.00 output per million tokens. For Claude Opus 5 from the same provider, $5.00 against $25.00. The ratio differs per provider and per model; the comparison puts them side by side.

For your costs this means that a short question with a long answer is more expensive than a long question with a short answer, even if the same number of tokens goes back and forth. If you have a model write a lot (an article, a long code change), you mostly pay for output. If you have a model read a lot (summarising a document, searching a codebase), you mostly pay for input.

• Input is processed in one go, output token by token; that costs more.
• The ratio is often one to four or one to five; for Claude Fable 5.1, $10.00 against $50.00.
• Having it write a lot costs output; having it read a lot costs input.

What does a conversation, a document, a month of building cost?

Calculate with the prices of Claude Fable 5.1 on 5 September 2026 as an example, and with round token numbers as an illustration. Take a conversation in which you send 20,000 tokens in total and get 4,000 tokens back. The amount sent is high because with every turn you send the whole conversation so far: the model remembers nothing between two turns. The input costs 20,000 divided by a million times $10.00, so $0.20. The output costs 4,000 divided by a million times $50.00, also $0.20. Together $0.40.

Having a document of fifty pages summarised: about 40,000 input tokens and say 1,000 output tokens, so $0.40 plus $0.05 is $0.45. An agent that reads and rewrites large parts of your code per instruction easily uses several million tokens a day. Then it is tens to hundreds of dollars a month, depending on how much you build and how big your project is. That last figure is not an estimate this article can make for you; it is a number you read in your provider's usage overview.

The lesson from the sum: single conversations cost cents, documents cost tens of cents, and agents that read and write code cost money. So if you build with an agent, you do not look at the price of a conversation but at the price of a day.

• A conversation of 20,000 in and 4,000 out: $0.40 at $10.00 and $50.00 per million.
• Summarising a document of fifty pages: about $0.45.
• An agent that builds: millions of tokens a day; read the usage, do not estimate it.

Why is long context expensive?

Because a model remembers nothing between two turns. Every time you ask something in an ongoing conversation, the application sends the whole conversation so far again as input: your first question, the first answer, your second question, and so on. So the tenth message in a conversation does not cost the tokens of that message, but the tokens of all ten. A long conversation therefore becomes more expensive per turn, even if you keep asking short questions.

The same applies to documents and code. An agent that sends your whole project with every instruction to understand it pays for that whole project with every instruction. That is why agents are so expensive compared to conversations. And that is why it helps to split a large project, or to let the agent read only the files the instruction needs.

Some providers also charge a higher price per token once the input exceeds a certain length, and most offer a discount on input they saw recently (caching). Both are in the provider's pricing terms and not in the price per million, and both can change your bill considerably. The comparison adds a note for a model where that applies.

• Every turn sends the whole conversation again; the tenth message costs ten messages.
• An agent that sends the whole project pays for the whole project every time.
• A higher price above a length, a discount on repeated input: read the terms.

Which price traps do you only see on the bill?

Six, and all six occur. The tokenizer: the same text costs more tokens with one model than with another, so do not compare only the price per million. The long context: a conversation or project that grows, grows in price per turn. Introductory prices: a new model is sometimes cheap for a while and later becomes more expensive; check the date next to a price. Peak and off-peak rates: some providers charge less outside busy hours or for work that can wait.

Reasoning levels: many models can think longer before they answer, and those thinking steps are output tokens you pay for without seeing them. A model on the highest setting can cost a multiple of the same model on the lowest. And access: a model that runs through an intermediary sometimes has a different price there than at the provider itself.

The remedy is always the same: do not look at the price but at the bill. Set a usage limit and an alert with your provider, and after the first week read what a day of work cost you. That number is the only price that counts for you.

• Tokenizer, long context, introductory prices, peak and off-peak, reasoning levels, intermediaries.
• Thinking steps are output you pay for without seeing it.
• Usage limit, alert, and after a week read off the price per day.

How do you choose a model on price without losing quality?

By putting price and quality side by side, and not just one of the two. The comparison on this site does that. For fifteen models it puts the price per million input and output tokens next to the score on the Intelligence Index of Artificial Analysis, an independent measuring party that tests models on the same ten evaluations. On 5 September 2026 the index is at version 4.2; a score from an earlier version is not comparable, and the comparison says so.

The pattern you then see: the most expensive models score highest, but the difference in score between the most expensive one and a model that costs a fraction is often smaller than the difference in price. For most work (rewriting text, answering a question, a summary) a mid-range model is more than good enough. For the hardest work (finding a complicated bug in code, processing a long document precisely) the most expensive model pays for itself in fewer retries.

So use two models: a cheap one for ordinary work and an expensive one for difficult work, and switch deliberately. Most applications and agents let you set that. If you do everything on the most expensive model, you pay for quality you usually do not need. If you do everything on the cheapest, you pay in retries.

• Price next to quality, with an independent index; the comparison does that with source and date.
• The difference in score is often smaller than the difference in price.
• Two models: cheap for ordinary work, expensive for difficult work.

How do you book it, and what about VAT?

As business expenses, like any other subscription or usage for your business. The providers charge in dollars and without VAT (btw). The debit on your account is in euros at the exchange rate of that moment, and that amount is what you book. Keep the provider's invoice or usage overview, because that is your supporting document; a credit card statement alone is not.

The tax explanation on this page follows Dutch law and is written for Dutch sole proprietorships (eenmanszaak). VAT is the point where things often go wrong with AI services from abroad. A provider outside the Netherlands usually charges a Dutch business owner no Dutch VAT: the VAT is then reverse-charged to you (verlegd). You report it in your VAT return, in the section for services from abroad. In the same return you deduct it again as input VAT (voorbelasting) if you are entitled to that. On balance nothing changes, but the return does have to show it. If you do not give the provider a VAT identification number (btw-identificatienummer), it sometimes does charge VAT. That is foreign VAT, which you may not deduct as Dutch input VAT.

So make sure your VAT identification number is in your account with the provider, and book the invoices with the reverse charge noted. The term reverse-charged VAT (verlegde btw) on this site explains the section of the return. The article about the first VAT return names this as one of the mistakes that occur most often among makers.

• Business expenses, booked in euros, with the provider's invoice as proof.
• Foreign provider: VAT reverse-charged, report it and deduct it in the same return.
• VAT identification number in your account, otherwise you pay foreign VAT you do not get back.

Sources

Checked on 5 September 2026. The prices come from the comparison on this site, which names the source and the check date per value; the token numbers in the sums are round illustrations. The rule of thumb of four characters per token is the providers' own and applies to English.

• basestep, the API comparison (every value with source and check date)
• Artificial Analysis, the Intelligence Index

Further reading

Three places that connect to this.