Model Implementation May 19, 2026

LLM pricing runs on the alphabet

Tokenizers charge different prices for the same meaning depending on language and phrasing, often before the model does any real work.

DL

Daniel Lopez

May 19, 2026 · 6 min read

Every time you send a prompt in a language other than English, the model spends more tokens processing it than it would in English. That gap is tokenizer arithmetic showing up before the model reasons.

The bill arrives in how your text is split into pieces before the model judges answer quality.

What a token is and why it matters

LLMs read tokens: variable-length chunks the model learns during training. English has efficient tokens; common words often fit in one. In other languages, the same word can require two, three, or more.

You pay more to say the same thing.

The tokenizer is an invisible tax. Your native language changes the multiplier.

Here are the real multipliers against English (baseline = 1.0×):

LanguageOpenAIAnthropicNotes
English1.0×1.0×baseline
Spanish1.2×1.62×second best
French1.79×notable
Chinese1.15×1.71×worse than expected
Arabic1.31×~3.0×severe penalty
Hindi1.37×3.24×severe penalty

Anthropic penalizes more than OpenAI across nearly every non-English language. Their tokenizer was built with heavy English bias, which is a training oversight rather than a pricing conspiracy.

Spanish is the exception that proves the rule

Spanish has the best multiplier of all non-English languages on Western models. 1.2× on OpenAI, 1.62× on Anthropic.

High training data volume plus structural similarity to English helps. Spanish still costs more than English. If you must run non-English production traffic, Spanish is usually the least bad option on Western tokenizers.

The honest counterargument

Weak English can cost more in rework than fluent Spanish saves in tokens. The real variable is tokens × output quality. Measure both.

The right move is to understand the trade-off and choose consciously, including when "always write in English" is wrong for your team.

Second order: caveman mode

Minimal phrasing (drop filler, connectors, padding) can shrink a 70-token answer toward 20. That discount is independent of language.

Caveman English still beats caveman Spanish on token count. Both beat verbose prose in high-volume systems.

What to do with this

  • Production prompts in high-volume systems: write in English if you can, or measure the actual cost before assuming language choice is free.
  • Personal use and exploration: any language is fine; know the bill.
  • Long prompts in high-penalty languages like Arabic, Hindi, or Russian: consider English for the prompt body, or evaluate models with better tokenizers for those languages (Gemini has reported advantages for Chinese, for example).

The tokenizer will not change tomorrow. Your prompt strategy can.

Bring us the AI decision you're trying to make.

We'll help you see the parts clearly.

Start a conversation