Why Speaking French to AI Costs More Than English

Using artificial intelligence in French can consume more tokens than asking exactly the same question in English. A study by Thomas Serval, published in September 2026, measured a 31% to 58% overhead for French, depending on the tokenizer. This invisible difference can increase API bills, reduce how much conversation history fits in memory, and change the conditions under which a model generates its answers.

We have grown accustomed to treating ChatGPT as a universal conversational partner. We speak our own language, it replies, and the illusion works well enough that few people wonder what happens between typing a sentence and the machine interpreting it. Yet languages are not equal within this machinery. Some fit almost naturally into the system. Others must pay an extra toll.

Why does a French sentence use more tokens?

Artificial intelligence does not read sentences as we do. Before a language model can process them, a component called a tokenizer splits text into units known as tokens. These units may be complete words, parts of words, or individual characters.

OpenAI offers a rough rule of thumb for everyday English: one token represents about four characters, or roughly three-quarters of a word. But this approximation does not apply equally to every language.

The reason is technical. Tokenizers learn to split sequences efficiently based on the material encountered during their design. When that material is dominated by English, many English sequences become easily recognized units. Equivalent French expressions may be split into several pieces.

A common English word may therefore fit into a single token, while its French equivalent sometimes requires two or three. The effect can be even stronger with certain non-Latin writing systems.

However, distinguishing high-resource languages from less represented ones is not enough. Reporting by Rest of World on research from the Coalition for Language Equity in AI Research emphasizes that the writing system itself has a major influence on fragmentation. Arabic and Hindi can therefore face penalties that cannot be explained simply by their number of speakers.

How much does this language tax really cost?

The figures are significant enough to be more than a technical curiosity.

According to Thomas Serval of NEOMA Business School, whose study used the multilingual NTREX-128 corpus, French requires 1.31 to 1.58 times as many tokens as English, depending on the system examined. The analysis compares seven tokenizers associated with OpenAI, Meta, Alibaba, DeepSeek, Google, Mistral, and Anthropic.

Another analysis, published by Prompt Inspiration in May 2026, estimates an average overhead of around 48% for prompts written in French. That estimate is broadly consistent with the earlier study, but the calculation methods are not necessarily identical. It would therefore be misleading to treat this percentage as a constant across all conversations and models.

The differences can be much larger elsewhere. According to Coalition for Language Equity in AI Research findings reported by Rest of World, the gap between the least and most expensive languages among the twenty most widely used languages online ranges from a factor of 1.5 to 5. When the analysis expands to two hundred languages, the gap can reach a factor of 17.

What do these numbers mean for your bill?

Consider a purely illustrative example. If processing a request in English costs €10 in tokens, a 48% increase would bring the equivalent cost to €14.80, assuming the same price per token. For a service handling many requests every day, that difference recurs with each use.

However, two situations must be distinguished. Someone paying a fixed ChatGPT subscription does not automatically receive a higher bill simply because they write in French. By contrast, a developer or company using an API billed per token may indeed bear this additional consumption cost.

The real issue is therefore not just paying more. It is getting a different amount of processing capacity for the same budget or token limit.

Why can this difference affect answer quality?

Every model has a context window: a limited number of tokens it can consider while working. That window includes instructions, previous messages, and supplied documents.

If two conversations express the same content but the French version requires more tokens, they take up different amounts of space in that window. The French conversation may reach its limits sooner.

Imagine a virtual private tutor. It supports a student and remembers their mistakes, difficulties, and explanations already provided. If this history uses more tokens in the student’s language, the service has less room to retain information useful for ongoing support.

This is precisely the example raised by Pratyush Choudhury, co-founder of Activate AI, in the Rest of World investigation. The language penalty can then become an educational disadvantage.

One essential distinction matters, however: using more tokens does not automatically mean the model reasons less effectively. Fragmentation puts more pressure on the context window and may increase latency, but on its own it does not demonstrate a systematic decline in the intellectual quality of responses.

Nevertheless, an academic study presented at EMNLP in 2023, examining the costs and usefulness of the OpenAI API across twenty-two languages, documents a broader inequality: many users of languages other than English face higher costs while receiving less favorable performance.

In other words, two distinct problems can compound each other: how sentences are split up, and how well the model can work in each language. Confusing them would make it harder to understand what actually needs improvement.

Are all AI models equally unfavorable to French?

No. The choice of tokenizer substantially changes the language-related cost.

In Thomas Serval’s comparison, Mistral Tekken appears to impose the smallest penalty on French among the seven systems studied. This suggests that some of the inequality can be addressed directly through technical design.

A compilation published by Prompt Inspiration also assigns Gemini 3.1 and Qwen 3.6 an overall overhead factor of 1.17 compared with English across the languages evaluated. Conversely, it reports an average factor of 2.07 for Anthropic and a ratio of 3.11 for a Hindi prompt using its API.

These comparisons should be treated cautiously. The latter figures come from a secondary compilation and are not enough to establish a universal model ranking. Above all, a token-efficient tokenizer is not necessarily paired with the model that gives the best answers. Response quality, reasoning ability, and token consumption must be assessed separately.

Can this language penalty be eliminated?

The encouraging answer is that the extra cost is not a mathematical inevitability. It stems partly from technical choices.

Thomas Serval’s study presents a tokenizer specifically adapted to French that reduces token consumption by 11.5% compared with the standard systems examined. The prototype also performs less well in certain situations. It is therefore no magic solution, but it offers concrete evidence that a different design is possible.

Another approach is to improve language representation in the resources used to develop these systems. The RESPIN project, mentioned by Rest of World, collects text and speech data from native speakers of thirty-eight dialects.

Users can already adopt several habits to limit the impact:

Compare tokenizers when API costs matter, rather than choosing solely on the advertised price per million tokens.
Monitor conversation history length in long chats or assistants handling large numbers of documents.
Test different phrasings with a visualization tool such as OpenAI Tokenizer to see how text is split.
Evaluate models in the language actually used, rather than relying only on English-language benchmarks.

Systematically translating requests into English might sometimes save tokens, but it is not a universal answer. It adds a step, can alter the nuances of the message, and does not guarantee a better final response.

The problem goes far beyond ChatGPT pricing

This technical bias reveals a deeper question: why should one language have less processing capacity than another to express the same ideas?

In traditional software, changing the interface language generally does not reduce the amount of work a user can accomplish. With language models, that equivalence is no longer guaranteed. Language becomes an economic and technical variable, even when it remains invisible on screen.

French speakers are not necessarily the most disadvantaged. Differences observed for certain regional and overseas languages, which can reach 3.3 times the English token consumption in observations surrounding Thomas Serval’s study, show that the issue extends beyond major international languages.

In my view, real progress will not simply mean making models more intelligent. It will also mean ensuring that this intelligence does not systematically cost more depending on the language used to address it. An AI claiming to be universal should not impose an invisible toll on people who do not speak English.

Frequently asked questions

Why does the same text in French or Arabic use more tokens than in English?

Tokenizers recognize certain sequences more efficiently when those sequences are common in their training material, which is often dominated by English. French and languages with other writing systems may require more fragments to represent equivalent content. The size of the gap depends on the tokenizer and language.

Does ChatGPT cost more when you write in French?

A fixed subscription does not automatically become more expensive for a French-speaking user. However, APIs billed per token may cost more: Thomas Serval’s study measures 31% to 58% more tokens for French, depending on the tokenizers compared. This consumption can also affect context limits.

Does ChatGPT reason less effectively in French?

Higher token consumption does not directly prove weaker reasoning. It can, however, reduce the room available for conversation history and documents, potentially affecting some lengthy tasks. The study published at EMNLP in 2023 also documents differences in performance and costs across languages.

Which models best reduce the language tax?

In Thomas Serval’s study, which focuses in part on French, Mistral Tekken has the lowest overhead among the seven tokenizers examined. Other comparisons highlight Gemini and Qwen on multilingual measures. Results depend on the corpus and tokenizer, however, so response quality should also be compared before choosing a model.