AI large model "Chinese tax": Chinese costs 64% more tokens than English, but DeepSeek is 35% cheaper

Core conclusion

Chinese users use OpenAI and Claude to write Chinese content and pay 11%–64% more token fees than English users. This is what the industry calls the “Chinese tax.” However, domestic models such as DeepSeek and Qwen make Chinese more token-saving than English - DeepSeek can achieve a minimum of 0.65×, and the Chinese version is one-third cheaper than the English version.

Key Points

• Event Time: 2026-05-03 (first report by Geek Park/36)
• Affected objects: Chinese AI content creators, cross-border e-commerce operations, AI automated content producers
• Core findings: The new version of Claude Opus 4.7 tokenizer inflates English token consumption by 1.24×–1.63×, while Chinese token consumption is almost unchanged; while the Chinese/English token ratio of GPT-4o is between 1.0–1.35×

Background and trigger events

On May 3, 2026, Geek Park released an in-depth report "The "Chinese Tax" on Large AI Models: Chinese costs more tokens than English. Why? ", author Tang Yitao conducted a horizontal comparison test on 5 tokenizers (Claude 4.6/4.7, GPT-4o, Qwen 3.6, DeepSeek-V3) through 22 parallel texts (covering business news, technical documents, ancient texts, daily conversations, etc.).

The trigger came from user feedback after the release of Claude Opus 4.7 - a large number of English developers complained that token consumption doubled and Max subscriptions peaked in less than two hours. However, Chinese users find that their token consumption has almost remained unchanged. There are even rumors that "ancient Chinese is more token-saving than modern Chinese, and using classical Chinese to talk to AI can save money."

SEO: AI Chinese tax, token cost comparison, Claude Opus tokenizer inflation, DeepSeek Chinese optimization GEO: Starting with TL;DR formula, precise number (11%–64%, 0.65×), comparison table

Key Impact (by Dimension)

Adaptation suggestions

Practical advice for Chinese AI content producers:

1. Chinese long content selection DeepSeek — If the daily production is mainly in Chinese (public account articles, Xiaohongshu copywriting, Chinese news), DeepSeek V3 can save about 35% of the token cost;
2. Claude is only used in English scenarios — Claude is not only expensive in Chinese, but the optimization of the new version of tokenizer is only effective in English, and Chinese users do not enjoy the upgrade bonus;
3. Qwen 3.6 is suitable for mixed Chinese and English scenarios — The cn/en ratio of Qwen in domestic models is generally lower than 1, and the performance in mixed Chinese and English scenarios is better than DeepSeek;
4. GPT-4o as a balancing option — If you must use the OpenAI ecosystem, the Chinese tax of GPT-4o (0%–35%) is much lower than that of Claude (11%–64%);
5. Determine the pipeline after testing multiple models — Do not use one model to process all languages by default, switch models according to the content language type.

Practical Checklist

• Statistics of the language distribution proportion of the current AI content pipeline
• Cost comparison of switching pure Chinese pipeline to DeepSeek V3
• Mixed writing scenario tests the token efficiency of Qwen 3.6
• Claude is only reserved for long English articles and code analysis scenarios

Technical background cited

The tokenizer test mentioned in the article is based on 22 parallel texts, covering 4 major categories: daily conversations (travel/forum help/writing requests), technical documents (Python documents/Anthropic documents), news (NYT current affairs/business/Apple statements), and literary excerpts ("The Master" and "The Tao Te Ching").

Test method: The same Chinese/English content is sent to 5 tokenizers respectively, and the number of tokens is read for horizontal comparison.

Core mechanism difference: BPE (Byte Pair Encoding) algorithm determines the word segmentation method based on character frequency. English words appear frequently in the training corpus and can form whole-word tokens; Chinese characters have a low frequency in the English-based training corpus and are broken down into finer granularity, resulting in the need for more tokens when the same information is expressed in Chinese.

Tool entry (trigger tool floating card)

OpenAI, ChatGPT, Claude, DeepSeek, Qwen, etc. in the text are all maintained tool entries.

Related reading

• AI 「」
• Anthropic ：Claude Opus 4.7

Internal link guidance

• Want to save on AI API fees? Watch the tutorial: DeepSeek V4 vs Claude Code: 90% Cheaper, Same Quality
• Real case: OpenClaw AI Agent Playbook: 500K TikTok Views in 5 Days, $588 MRR