What an AI token is and what it really costs

An AI token is a fragment of text, most often a piece of a word, that AI models use to read and write. It's also the unit they bill in. When a price list says "$3 per million tokens", it states the price of the raw material of text, the way a baker might sell flour by the kilo whatever recipe you make from it.

Two questions usually come together, what exactly a token is and how much it costs. The answers rest on concrete equivalents such as an email, a page or a book. They also cover languages like French, which use more tokens than English, along with today's official price lists provider by provider.

01Key takeaways

• A token is a fragment of a word. Count about 0.75 word per token in English and around 0.5 to 0.6 in French.
• Prices are shown per million tokens, with input and output listed separately. Output costs two to six times more than input depending on the model.
• For scale, an email weighs a few hundred tokens, a blog article 2,500 and a whole book about 130,000 (figures for French text).
• French and most other languages need more tokens than English for the same content, so texts in those languages cost a little more than English examples suggest.
• At these rates a 1,500-word article costs a few cents. The token is a tiny unit with a tiny price.

02A token in one sentence

A token is the unit a language model uses to cut up text. Before reading your sentence, the model slices it into fragments taken from a fixed vocabulary of a few tens of thousands of pieces, learned once and for all during training. Common words fit in one token, while rare or long words split into several. Punctuation and spaces count too. The model never sees your letters. It sees a sequence of fragment numbers, which it handles and extends before turning them back into text.

You don't need to be an engineer to care, because everything is measured in tokens. That includes the maximum length of a conversation, the generation speed and above all the bill. Once you understand this unit, you can read every price list on the market. It's the first step in the entry point on pay-per-use AI, where the rest of the topic unfolds.

03Token, word, character and what each one counts

Three units coexist and often get mixed up. The character is the letter, the word is the human unit and the token is the machine unit, sitting between the two. The French greeting "Bonjour" fits in one or two tokens depending on the model. The very long "anticonstitutionnellement" splits into several fragments. A short word like "le" takes a single one, since the space before it usually belongs to the same fragment.

The practical rule to remember is that one token equals about three quarters of a word in English and a little more than half a word in French. The other way round, a thousand English words come to roughly 1,300 tokens and a thousand French words to roughly 1,700 to 2,000. These two ratios cover every everyday calculation. When you need the exact count, the providers' online tokenizers give it.

04Tokenization without the jargon

The machine decides where to cut by statistics. While the model is being prepared, an algorithm (the best known is called BPE) studies huge volumes of text and builds a vocabulary of the most frequent fragments, typically 30,000 to 50,000 entries. Very common letter sequences become tokens in their own right. Everything else is assembled from smaller fragments.

That has two concrete consequences. Texts full of rare words, jargon or proper names use more tokens than ordinary prose, because the vocabulary doesn't know them as a single block. Each model family also has its own way of cutting, so the same text doesn't weigh exactly the same at two providers, although the gaps stay modest on the scale of a bill. There's no point choosing a provider on that basis. The quality produced and the price per million weigh far more.

05Tokenization by example

Take the sentence "The bakers make great pastries". It comes to five or six tokens for five words, since each common English word has its own fragment. Run the same exercise on the French version of the sentence and the count climbs. Its opening word "Les" stays whole and "boulangers" splits into two or three pieces. The verb takes one or two tokens, the word for excellent splits around its apostrophe and "viennoiseries", a long and rare word, breaks into several fragments. The total is about ten tokens for six words, the usual French ratio.

That makes the extra load of French visible with no theory at all. The providers' online tokenizers let you play this way with your own sentences. It's five instructive minutes that teach more than any diagram ever drawn on the subject of tokenization.

06Why French and other languages cost a bit more

Tokenization vocabularies were historically built on bodies of text dominated by English. English fragments are better represented in them and therefore more economical. For the same content, French needs more tokens than English and languages written in ideograms usually need even more. You can see the gap yourself by pasting the same text in two languages into a provider's token counter.

At current prices there's little cause for worry. The extra load turns a one-cent article into a cent-and-a-half article. It's a factor to know so you read English-language examples correctly when you work in French, never a reason to change your working language. Recent models are also narrowing the gap as their vocabularies diversify.

07How many tokens your content holds

Indicative equivalents for everyday text, ratios observed on June 13, 2026

Content	Words	Tokens in English (about 0.75 word per token)	Tokens in French (order of magnitude)
Ordinary email	150	~200	250 to 300
Typical web page	500	~670	850 to 1,000
Full blog article	1,500	~2,000	2,500 to 3,000
Thesis or long report	20,000	~27,000	~35,000
Whole book	80,000	~107,000	~130,000 to 160,000

Keep the middle line in mind. The blog article at about 2,500 output tokens is the yardstick used in every cost calculation on the site, measured on French text. In English the same article comes closer to 2,000 tokens, so English estimates come out a little cheaper. The million tokens of the price lists represents about 750,000 words, according to OpenAI's official benchmark. That's the equivalent of ten novels, far more than a month of ordinary work.

08Where this odd way of cutting comes from

Models don't simply read words because there are too many of them. Between languages, conjugations, typos and invented words, no finite dictionary could cover everything. Fragments solve the problem. With a few tens of thousands of well-chosen pieces, you can rebuild any text in any language, unknown words included.

Reading character by character would have worked too, but at the cost of endless sequences that are expensive to process. The token is the compromise, big enough to be efficient and small enough to be universal. This engineering history explains everything else, from the ratios per language to the price per million tokens. Once you know it, nothing in the price lists looks strange anymore. The same goes for the way models behave with invented words.

09Why the token is the billing unit

Providers bill per token because that's their real cost. Each fragment read or written uses computing power on their machines, in proportion. The unit is therefore honest by design, finer than a subscription and more neutral than an in-house credit whose value depends on who sells it. Above all, it lets you compare offers, since every price list speaks the same language.

That's exactly what makes BYOK easy to read. Your API key buys tokens at the public rate, the statement counts the tokens used and you can check the multiplication yourself. No other formula on the market gives that end-to-end traceability, from the word fragment to the invoice line. In-house credits do the opposite and convert your tokens into a private currency whose exchange rate you don't control. Nothing about that is illegal, but comparison becomes impossible and that's rarely an accident.

10What a million tokens costs on today's price lists

Prices per million tokens recorded on October 16, 2026 from the official price lists cited in the sources

Model	Input	Output
Mistral Small 4	$0.15	$0.60
DeepSeek V4.1-Flash (peak hours)	$0.30	$1.20
GPT-5.4-mini	$0.75	$4.50
Claude Haiku 4.5	$1	$5
GPT-5.4	$2.50	$15
Claude Sonnet 4.6	$3	$15

Put simply, an article with 2,500 output tokens costs a third of a cent at most on budget models and about four cents on the high end. The full comparison, flagship models and special cases included, sits in the provider price comparison. Above all, remember the order of magnitude, which is cents everywhere you look.

11Input versus output, the gap to know

Price lists show two columns because the model bills two jobs. It reads your request (the input), then writes the answer (the output). Output costs two to six times more depending on the model, because generating takes more computing than reading. In most creative uses your briefs therefore weigh less than the texts produced. The reverse holds for document work, where you send a lot to get back a short summary.

Two levers follow from this gap and each has its own guide. Writing shorter instructions lightens the input, while prompt caching explained covers how repeated input gets billed at a fraction of the price. The full detail of the gap, with worked examples, is in the input versus output breakdown.

12Three common misunderstandings, cleared up

The first belief is that one token equals one word. That rounding skews every calculation, because the output of a 1,500-word article weighs 2,500 to 3,000 tokens in French, far from 1,500. The mistake makes you underestimate usage by a good third. With the ratios given earlier it disappears for good and your estimates land where they should.

The second belief is that spaces and punctuation don't count. They do, since every character sent or produced belongs to a billed fragment, formatting and line breaks included. The effect is marginal on an article. It becomes noticeable on highly structured data such as tables or code, which hold more symbols than words.

The third belief is that the more tokens a text uses, the better it is. There's no link at all, because the token count measures length and never quality. A wordy model costs more without writing better. Instructions that ask for brevity win on both fronts, readability and the bill, a double dividend that's rare in optimization.

13How counting in tokens changes your habits

Start with lean instructions. Once you know every line of the brief is billed as input, you stop pasting three pages of context for a two-sentence rewrite. The reflex settles in within a week and costs no comfort. It often improves the answers too, since clear instructions guide better than buried ones.

Reading the statement helps just as much. Every provider shows usage in tokens. Knowing that an article is worth 2,500 turns those columns of figures into useful information. You spot the greedy automation or the overly wordy template at a glance, things you'd never see otherwise. Two minutes with the monthly statement replace all the guesswork.

Comparing models calmly completes the set. Price gaps per million tokens look huge, from twenty to twenty-five times between the cheapest and the dearest in the table. Brought back to one article, they separate a sixth of a cent from four cents. Once you understand the unit, the price lists lose their drama and the choice comes back to what counts, the quality produced for your use.

14The context window, the other number in tokens

The token also measures a model's working memory. That's its context window, the amount of text it can consider at once, request and answer included. Common models list it in hundreds of thousands of tokens and the largest in millions, enough to take in whole reports or even books in a single request. Watch the budget side, though. Filling a large window is paid in input tokens at every call, which explains the bills of conversational use with very long context.

For a content creator the consequence is simple. The window is almost never the constraint, because a whole article fills only a fraction of it. It matters for heavy document work such as corpus analysis or case-file summaries. There, checking the window of the chosen model before you start avoids splitting the work for nothing and getting truncated results with no explanation.

15From the concept to the bill

With the concept in hand, two next steps make sense. To put a figure on your own uses, the method to price a call applies these ratios to your real volumes in ten minutes. For the full market overview, the real cost of AI in figures lines up all the dated price lists. When spending becomes a concern, every way to cut costs adds up without changing your habits.

To try it with no commitment, you can get a key without paying from several providers, paste it into a tool and watch the token counter run on real content. With OpenTahr, the free account bills nothing per generation, since each one is paid in tokens at your provider's cost price. You can create a key step by step in five minutes, with one guide per provider.

16Frequently asked questions

How many words is a token?

A token is about three quarters of a word in English and a little more than half a word in everyday French. The other way round, a thousand English words come to roughly 1,300 tokens and a thousand French words to roughly 1,700 to 2,000. The providers' online tokenizers give the exact count for a specific text when you need it.

How much does a token cost?

A single token costs a tiny fraction of a cent. Prices are shown per million tokens, from fifteen cents on budget models to $15 on the most expensive models in the table, with input and output listed separately. A full blog article therefore costs cents at today's official price lists, which are cited in the sources.

Why does French use more tokens than English?

Tokenization vocabularies were built on bodies of text that were mostly in English, so English fragments are more economical. The extra load varies with the text and stays harmless at current prices. It's shrinking gradually as the vocabularies of new models become more multilingual.

How do I count the tokens in my own texts?

You have three options. The rule of three with the ratios given earlier is enough day to day. Your provider's online tokenizer handles a specific text, while the API statement counts every call after the fact, down to the token. Most BYOK tools also show usage as you go, request by request, which makes learning the orders of magnitude almost automatic.

Do other languages use another word for token?

French officially translates it as jeton, but token is the term used in the field and in the documentation. The site sticks with token because it's the word you'll see on the price lists. Anyone you talk to will understand both terms without the slightest difficulty.

Transparency note. Prices were recorded on October 16, 2026 from the providers' official price lists, cited in the sources. The word-to-token ratios are observed orders of magnitude that vary with texts and models. The benchmark of 1 million tokens for about 750,000 words comes from OpenAI's official page. None of the providers named here pays us.

17Sources

1. OpenAI, official API price list (input and output tokens)
2. Anthropic, official Claude API price list
3. Mistral, official API price list
4. Mistral, API price list per model (Small 4 at $0.15 and $0.60)
5. DeepSeek, official API price list (peak and off-peak rates)