Hugging Face's logo

Datasets:
JOHNISKING
/
token-cost-by-programming-language

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
code
programming-languages
tokenizer
llm-cost
claude
gemini
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
token-cost-by-programming-language / README.md
JOHNISKING's picture
JOHNISKING
Add Zenodo DOI
27fb9f8 verified
preview code
|
Raw History Blame Contribute Delete
4.12 kB
metadata 
license: cc-by-4.0
language:
  - en
pretty_name: Token cost by programming language (Claude, Gemini, GPT)
size_categories:
  - n<1K
tags:
  - code
  - programming-languages
  - tokenizer
  - llm-cost
  - claude
  - gemini
  - gpt
configs:
  - config_name: default
    data_files:
      - split: train
        path: token-cost-by-programming-language.csv

Token cost by programming language: the same programs in 10 languages on Claude, Gemini and GPT

Two small programs (an inventory class and Dijkstra's shortest path) written in Python, Ruby, JavaScript, TypeScript, PHP, Rust, Java, C++, Go and C, with input tokens counted on Claude Opus/Sonnet 5.5, Claude Haiku 4.5, Gemini 3.x and GPT (o200k_base).

Headline: both programs together cost 809 GPT tokens in Python and 1,536 in C (+90%). Python and Ruby were the cheapest on every tokenizer; Go and C the most expensive. Claude 5.5 needed 43% to 59% more tokens than GPT for the same code.

Key findings

  • Python and Ruby are the cheapest: 809 and 819 GPT tokens. JavaScript needs +20%, Java +50%, Go +68% and C +90%.
  • The ranking holds on every tokenizer tested.
  • Claude 5.5 costs 43% to 59% more than GPT for the same code (largest gap: Java). Gemini needs 12% to 25% more than GPT.
  • Types add tokens: TypeScript needs 11% more GPT tokens than the same JavaScript.
  • File size is a rough guide only: Java has the most characters, but Go and C need more tokens.

Files

  • token-cost-by-programming-language.csv: one row per language and program (inventory, dijkstra and total) with language, program, characters, lines, claude55_tokens, claude45_tokens, gemini_tokens, gpt_tokens
  • token-cost-by-programming-language.json: the same data plus all 20 source files, so every count can be reproduced

Data (both programs together)

Language Characters Claude Opus/Sonnet 5.5 Claude Haiku 4.5 Gemini 3.x GPT (o200k) GPT vs Python
Python 2,887 1,176 970 963 809 ±0%
Ruby 2,478 1,181 961 944 819 +1%
JavaScript 3,195 1,442 1,199 1,184 967 +20%
TypeScript 3,610 1,590 1,311 1,288 1,070 +32%
PHP 3,697 1,609 1,311 1,262 1,128 +39%
Rust 4,081 1,815 1,452 1,411 1,189 +47%
Java 5,101 1,934 1,551 1,511 1,217 +50%
C++ 4,793 1,933 1,626 1,617 1,343 +66%
Go 4,292 2,143 1,749 1,701 1,359 +68%
C 4,843 2,219 1,859 1,820 1,536 +90%

Method

  1. Each program written idiomatically in each language with no comments. Every version was compiled or run and printed exactly the same output as the Python version.
  2. Claude: each file sent through Claude Code (claude -p) with a fixed one-line system prompt; input tokens minus the same run with a one-character message. Opus 5.5 and Sonnet 5.5 gave identical counts; a re-run gave identical counts.
  3. Gemini: Google AI Studio prompt token counter on Gemini 3.8 Flash, minus its one start token.
  4. GPT: tiktoken with o200k_base.

Limitations

  • Two small programs. Comments, long names or boilerplate in real projects can shift the numbers; the ranking is more stable than the exact percentages.
  • Coding style changes the count within a language.
  • Input tokens only; not a measure of how well each model writes each language.

Citation

TokenSave (2026). Token cost by programming language: the same programs in 10 languages on Claude, Gemini and GPT [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23446563

Count tokens for your own code at tokensave.app.