ModelVerdict
EnglishEN

The language tax on tokens

A cheaper price list doesn't mean cheaper to run

The same text in Czech, German, French, Spanish or Polish takes more tokens than in English, and tokens are what you pay for. Every model has its own tokenizer, so the surcharge differs between models.

Same text, Claude Sonnet 5

English100 tokens
Czech152 tokens

How many more tokens Czech takes

average over 50 paragraphs
  • Claude Sonnet 5×1.52 · +52 %
  • Gemini 3.8 Flash×1.64 · +64 %
  • Gemini 3.5 Flash Lite×1.64 · +64 %
  • Gemma 4 31B×1.64 · +64 %
  • Qwen3.8 Flash×1.66 · +66 %
  • GPT-6 Luna×1.71 · +71 %
  • GPT-6 Sol×1.71 · +71 %
  • Mistral Medium 3.5×1.73 · +73 %
  • DeepSeek V4.1 Flash×1.85 · +85 %

English baselinesurcharge

DeepSeek V4.1 Flash uses 22 % more tokens for the same text than Claude Sonnet 5.

Real cost calculator

What the same documents cost per month in two languages, English by default. Measured where we have documents in the language; elsewhere estimated from the language tax (≈).

Task
SummariesDocument Q&A
Languages→
Documents per month
ModelDifferenceMonthly cost

Monthly cost at the selected volume, converted at the rate in the footer. Excluding VAT. ≈ = estimate: the surcharge measured on real documents in another language, scaled by this language's language tax. Accuracy is only shown for languages we test documents in.

All languages and models
ModelCzechGermanFrenchSpanishPolishEnglish
Claude Sonnet 5×1.52×1.88×1.53×1.50×1.76×1.00
DeepSeek V4.1 Flash×1.85×1.59×1.58×1.54×1.86×1.00
GPT-6 Luna×1.71×1.29×1.34×1.31×1.77×1.00
GPT-6 Sol×1.71×1.29×1.34×1.31×1.77×1.00
Gemini 3.5 Flash Lite×1.64×1.33×1.38×1.30×1.60×1.00
Gemini 3.8 Flash×1.64×1.32×1.38×1.30×1.60×1.00
Gemma 4 31B×1.64×1.33×1.38×1.30×1.60×1.00
Mistral Medium 3.5×1.73×1.39×1.36×1.38×1.77×1.00
Mistral Small 4——————
Qwen3.8 Flash×1.66×1.28×1.37×1.32×1.59×1.00