Can conversing in classical Chinese with AI save tokens? A screenshot sparked a discussion: Engineer: Actually, using English is the way to go.

“Can using Classical Chinese reduce tokens?” — This question, accompanied by a screenshot of Claude Haiku 4.5 replying in Classical Chinese, “Indeed. Classical Chinese is more concise, with refined vocabulary, which can certainly reduce token consumption,” sparked a serious and interesting technical discussion in the community.

The logic behind this question

Intuitively, Classical Chinese is more succinct than vernacular Chinese — the character “蝶” stands for “butterfly” in vernacular; the character “可” means “can.” If each Chinese character counts as one token, theoretically, Classical Chinese could indeed save tokens. Grok also confirmed this assertion in the discussion thread by responding in Classical Chinese.

Engineers’ rebuttal: The tokenizer is key

However, several engineers pointed out a technical detail often overlooked — tokens do not equate to character count. The tokenizers of Western models like OpenAI are optimized for English; when processing Chinese, a single character often requires 1-2 tokens, and traditional characters sometimes consume even more tokens than simplified ones. In other words, “可” and “可以” might both be 2 tokens in certain models, meaning fewer characters does not necessarily mean fewer tokens.

The conclusion after testing is: American models are most economical in English, while Chinese models are most economical in modern Chinese; using domestic models, the token cost for Chinese content can be about 20% cheaper than in English.

Another unexpected finding: Classical Chinese might be easier to “jailbreak”

A more interesting observation emerged during the discussion — mainstream LLMs have hardly any safeguards against Classical Chinese, making it easier to bypass security restrictions and even extract content that the model would usually refuse to answer. It is reported that papers from ICML or ICLR documented this phenomenon.

Quality issues with Classical Chinese reasoning chains

Another rebuttal came from practical experience: “Using Classical Chinese reasoning chains can lead to a drop in quality. What a normal reasoning chain can answer correctly may result in errors with a Classical Chinese reasoning chain.” The logic is straightforward: LLM training data is primarily composed of modern English and modern Chinese, with Classical Chinese data comprising less than one-tenth; asking it to think in Classical Chinese is akin to making it reason in an unfamiliar language, thus naturally amplifying the hallucination rate.

Conclusion: It’s a good meme, but not a good engineering strategy

The outcome of this discussion is broadly: for Western models, using English is the true way to save tokens; for domestic models, modern Chinese is more stable than Classical Chinese. The “token-saving” effect of Classical Chinese is likely offset at the tokenizer level, potentially leading to a decline in reasoning quality. However, this screenshot did achieve another goal: it transformed a dull AI cost issue into an engaging discussion that everyone could participate in.

This article, “Can conversing in Classical Chinese and AI save tokens? A screenshot ignites discussion, engineers: actually, using English is the way,” first appeared in Chain News ABMedia.