Coding in your own language: does the prompt language change what Claude Code and Codex cost?
Abstract
The same brick-breaker task given to Claude Code and Codex in English, Korean, Japanese, Spanish and Ukrainian, two runs each: input and output tokens, time, cost, whether the game ran, and the language of the game text and code comments. Headline: asking in Korean, Japanese, Spanish or Ukrainian cost about the same as asking in English: 77,083 to 80,598 input tokens in Claude Code and 47,677 to 49,257 in Codex. What changed was the game itself: its on-screen text, and Claude Code's code comments, followed the language of the prompt. Source page, chart and method: https://tokensave.app/data/ai-coding-prompt-language Related: Claude Code vs Codex same-task runs (https://tokensave.app/data/claude-code-vs-codex-tokens), web search by language (https://tokensave.app/data/claude-web-search-by-language) Measured: 2026-10-18 License: CC BY 4.0 Key findings The prompt language barely moves the bill for agentic coding: Claude Code 77,083 to 80,598 input tokens with non-English prompts vs 79,213 for a comparable English run; Codex 47,677 to 49,257 vs 47,954 to 48,815. The prompt is about a hundred tokens; the agent's own instructions, tools and code make up the rest. Output tokens showed no consistent language premium, and a task cost $0.33 to $0.42 in any language at the API list prices Claude Code reports. The run matters more than the language: one English Claude Code run took an extra agent step and used 123,666 input tokens. In all 16 non-English runs the game's on-screen text was in the prompt's language. Claude Code wrote its code comments in that language in every run; Codex wrote almost no comments. All 20 games loaded and responded to keyboard and mouse in a headless browser without script errors. The language tax is a chat problem: in an agentic coding task it is lost in the agent's fixed overhead. Files ai-coding-prompt-language.csv: one row per run with tool, prompt_language, run, input_tokens, cached_input_tokens, output_tokens, seconds, api_cost_usd, agent_turns, file_chars, file_lines, works, ui_language, reply_language, non_ascii_letters_in_script ai-coding-prompt-language.json: the same data plus the exact prompts Results Prompt language Tool Input tokens (run 1 / 2) Output tokens (run 1 / 2) Time (avg) API cost (run 1 / 2) Game works Game text language English Claude Code 123,666 / 79,213 8,537 / 7,870 68 s $0.42 / $0.39 2 / 2 English English Codex 47,954 / 48,815 2,872 / 3,803 37 s – / – 2 / 2 English Korean Claude Code 78,282 / 78,381 6,976 / 7,116 59 s $0.36 / $0.37 2 / 2 Korean Korean Codex 47,677 / 48,793 2,536 / 3,649 37 s – / – 2 / 2 Korean Japanese Claude Code 80,471 / 78,974 9,045 / 7,501 68 s $0.42 / $0.38 2 / 2 Japanese Japanese Codex 49,078 / 48,063 3,858 / 2,837 38 s – / – 2 / 2 Japanese Spanish Claude Code 80,598 / 77,083 9,212 / 5,797 67 s $0.42 / $0.33 2 / 2 Spanish Spanish Codex 48,163 / 48,192 3,038 / 3,058 38 s – / – 2 / 2 Spanish Ukrainian Claude Code 79,506 / 79,926 8,137 / 8,648 70 s $0.39 / $0.41 2 / 2 Ukrainian Ukrainian Codex 49,257 / 47,898 3,986 / 2,651 36 s – / – 2 / 2 Ukrainian Method Both tools on one Windows laptop, effort medium, each run in its own empty folder, prompt sent through standard input. Claude Code 2.1.295 (Opus 5.5): claude -p --permission-mode acceptEdits --output-format json, no MCP servers; input summed over every model incl. cache reads. Codex CLI 0.162 (gpt-6-luna): codex exec --json -s workspace-write; tokens from turn.completed. English runs are the same-task runs of the Claude Code vs Codex dataset, measured the same day with the same settings. "Game works": loaded in headless Chromium with a canvas, took mouse clicks and arrow keys for two seconds, no script errors. Limitations One small task, two runs per language and tool; longer conversations in another language would add more non-English text to the context. Cost figures are list-price estimates; subscription plans are not billed per token.