The same question can cost more when put to an AI in French rather than English because the two versions are not necessarily split into the same number of tokens. In a comparison of 1,997 translated news sentences presented by Alain Goudey in October 2026, French versions used 1.31 to 1.58 times as many tokens as English versions, depending on the tokenizer. This gap can affect API bills and how much text a model can examine.

The difference is subtle: on screen, two requests look equivalent. Yet they may occupy different amounts of space in the system processing them. It is a surprising reminder that equality between languages concerns not only response quality but also how machines count the characters we give them.

Why can a French sentence take up more tokens?

A language model does not simply receive words separated by spaces. A tokenizer prepares text as units called tokens. Depending on how it splits text, a unit may be a whole word, part of a word, a punctuation mark or a space. An uncommon or invented term may be broken into several pieces.

This is not a universal grammatical rule. It depends on the model's tokenization tool. The same passage can yield different counts in two systems. Comparing only word counts in French and English is therefore like calculating a bill in the wrong units.

Language and writing system also matter. Rest of World points out that language rarity alone does not explain the differences. In research published in 2025, Teklehaymanot and Nejdl identified tokenization difficulties with certain non-Latin scripts and morphologically complex languages. Their findings are not a definitive ranking of every language.

How much can the difference cost?

For APIs that charge per token, distinguish the tokens sent to the model from those generated in its reply. A request requiring more input tokens can raise that portion of the bill at the same rate. But it cannot predict the total cost: response length and model-specific prices also matter.

Alain Goudey's comparison offers a concrete indication: across 1,997 translated sentences, French required 31% to 58% more tokens than English, depending on the tokenizer. This is neither an automatic surcharge imposed by providers nor a rule that applies to every prompt. It is a measured difference in a specific corpus.

On a broader scale, Rest of World reported in October 2026 gaps of 1.5 to 5 times between the extremes among twenty languages. When the comparison expanded to two hundred languages, the article noted a gap of up to seventeen times in one model. These ranges reveal an infrastructure inequality, not the cost of a French speaker's next conversation.

Why does it also reduce room for documents?

Price is only the visible part of the problem. A model has a limited context window, also measured in tokens. This space is shared by the current question, earlier messages, supplied documents and the answer to be generated. A token used in one place cannot be used elsewhere in the same window.

Consider a professional setting: a long conversation accompanies several documents to analyze. If the exchanges and documents consume more tokens, less capacity remains for other files or useful history. This is not just a budget issue; it affects what the tool can keep available when answering.

This gives concrete meaning to the phrase “access to AI tools.” Two people using the same model and context limit do not necessarily fit the same amount of usable content in their respective languages. It does not prove answers are always worse in one language; it shows that an identical token limit can mean different working capacity.

How can you measure the gap in your own texts?

Rather than seeking a universal percentage, run a reproducible test. Choose text representative of your work: an instruction, document excerpt or recurring question. Prepare French and English versions with as close a meaning as possible. Do not compare a detailed request with a summary.

• Choose the intended model. The tokenizer matters as much as the language; results from another model may mislead.
• Count both inputs separately. Compare tokens, not just words or characters.
• Estimate the output too. A long answer can change the calculation; apply the relevant input and output rates.
• Repeat with several typical texts. One example may not reflect everyday usage.

The Vantaige calculator lets you paste text, choose a provider and model, and obtain a token count and estimated API cost. It is a useful starting point, not a guarantee of the final bill: actual generated text may differ from a simulation.

One more precaution matters: do not compare tokenizers while ignoring prices. An APIYI test reported 1,630 output tokens with Gemini 3 Flash versus 812 with DeepSeek V3.2 for a French translation of the same Chinese article. This compares output counts across two models; it does not establish which is cheaper on its own.

Should you therefore write to AI in English?

Not automatically. If you work in French, a French instruction may be easier to check, correct and incorporate into the final document. English is worth testing when volumes or API costs make the gap significant, but do not assume it always wins. Compare the complete task, not an isolated sentence.

To me, the real issue is not pushing everyone toward a dominant language to save tokens. It is making costs transparent. Users should be able to see how much capacity their text consumes before choosing a model or preparing a large document. What can be measured can be debated; what stays invisible too easily looks like a natural limit.

Frequently asked questions

Why do equally long texts use different numbers of tokens?

Because a tokenizer does not necessarily split text word by word. Segments vary with words, language, script and model. Passages of similar visual length can therefore produce different counts.

Does French always cost more than English?

No, this is not an absolute rule. Alain Goudey's comparison of 1,997 translated sentences found extra tokens in French, but the amount varied by tokenizer. For a particular request, count both versions and apply the model's prices.

How does this affect document analysis?

A model's context window is limited in tokens and must accommodate messages, documents and the response. If the same content consumes more, less room remains for everything else. The exact effect depends on the model, text and conversation.

How can I check the cost of my own questions?

Write two versions with the same meaning and count their tokens using the intended model's tokenizer. A tool such as Vantaige can estimate counts and API costs. Separate input and generated tokens, and treat the result as an estimate, not a guaranteed bill.