Study: French costs up to 58% more tokens than English across 7 LLMs
A paper measures a persistent "language tax" in 2026 tokenizers, and tests a French-optimised alternative to Mistral's Tekken.
The facts
- French requires 31% to 58% more tokens than English across seven tokenizers tested: OpenAI o200k, Llama 3, Qwen3, DeepSeek V3/V4, Gemma 3, Mistral Tekken, and Claude generation-5 (via Anthropic's counting API)
- Simplified Chinese is cheaper than French on six of the seven tokenizers, ranging from 5% fewer to 40% more tokens than English
- Regional and overseas languages of France pay roughly 1.6 to 3.3 times the English token count
- The authors' prototype, Baracoda FR v1.2, uses 11.5% fewer tokens than Tekken on French and 3.7% fewer on English, but performs worse on other languages and does not beat CroissantLLM at comparable vocabulary size
Researcher Thomas Serval has put a number on something French-speaking developers have long suspected. Their language costs more to run through large language models. He tested seven tokenizers from widely used 2026 models on 124 reference translations (NTREX-128) and the Universal Declaration of Human Rights. French needs 31% to 58% more tokens than English to encode the same content. Regional and overseas languages of France fare worse still, paying 1.6 to 3.3 times the English token count.
The comparison that will sting most is with Chinese. Simplified Chinese, despite using a non-Latin script, is cheaper than French on six of the seven tokenizers. It ranges from 5% fewer tokens than English to 40% more. The paper argues this gap compounds in agentic use. Conversation history gets re-sent on every turn. Pricing tiers bill per token. Context windows are fixed in token counts, not characters. Together these turn a fixed percentage premium into a larger absolute cost and a smaller effective context for French users.
Serval ran a controlled experiment showing the fix is tractable but not free. Adding French text to tokenizer training data, using BPE on Europarl with a 50,000-token vocabulary, quickly cuts the premium. But returns diminish, and English efficiency pays a growing cost. The paper's own prototype, Baracoda FR v1.2, is built at Tekken's vocabulary size. On a final test set of six held-out corpora, with the testing protocol fixed in advance, it uses 11.5% fewer tokens than Tekken on French and 3.7% fewer on English. Yet it is worse on other languages. And it does not beat the existing French-tuned model CroissantLLM at a comparable vocabulary size.
The paper is explicit about its limits. These are segmentation counts, not evidence that a cheaper tokenizer changes model quality or end-to-end task cost. That gap, between tokens saved and dollars or accuracy saved, is where the real argument will be settled.
Why it matters
Anyone billing or budgeting in French, or building for French-speaking users, is paying a hidden surcharge baked into tokenizer vocabularies trained predominantly on English. That surcharge shrinks effective context windows as well as inflating API bills. The result argues for scrutinising tokenizer training data before assuming cross-lingual pricing parity. But the paper itself stops short of showing that a better tokenizer improves downstream quality or lowers real task cost.
Primary source
arXiv cs.AI — arxiv.orgWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.