Logographic languages look like they should compress agent prompts, but BPE tokenization fragments them into per-byte pieces. GLM-5 encodes Chinese at 0.98x the English token count, while MiniMax-2.7 inflates to 1.28x. The round-trip translation overhead (Chinese reasoning to English output) taxes every step of the agentic loop. The cost model and the empirics, side by side.
A Simplified Chinese character can encode an entire semantic concept in a single logogram, where English spreads that same concept across multiple characters, spaces, and grammatical particles; the arithmetic looks like it ought to be obvious, since if you ask an agentic LLM to conduct its chain-of-thought reasoning in Chinese, the intuition says the context window should then hold more reasoning per token. Practitioners call this “vibe coding,” and it is an appealing idea; I ran through the arithmetic myself when I was planning a multilingual agentic pipeline, and the tokenizer turns out not to cooperate in the way the intuition promises.
Modern LLMs do not process raw characters; they process tokens produced by Byte-Pair Encoding (BPE), a statistical compression algorithm trained predominantly on English-heavy internet corpora, and BPE builds its merge table from co-occurrence statistics in the training distribution, so it compresses frequent English words into single tokens efficiently while Chinese logograms, which are underrepresented in that distribution, get fragmented into multiple sub-word units or raw bytes.
Ahia et al. (2023) and Petrov et al. (2023) quantify the disparity directly; for many Western LLMs, encoding Simplified Chinese requires two to three times as many tokens as encoding the equivalent semantic content in English, so the character-level density that makes the hypothesis appealing tends to be consumed by the tokenizer before the model ever sees it.
The distinction worth holding onto is between morphosyntactic density (characters per concept, a property of the writing system) and tokenization density (tokens per concept, a property of the model’s vocabulary); Chinese wins the first metric, and English wins the second on most production models.
Whether Chinese saves tokens at all turns out to depend on the specific model’s vocabulary allocation; Ren et al. (2026) tested the vibe-coding hypothesis empirically across multiple models and found the efficiency to be entirely model-dependent:
The token count is not the only variable in play; Ren et al. also measured task resolution rates and found that models prompted to reason in Chinese frequently exhibited lower problem-solving rates than the same models prompted in English, so the metric that tends to govern a deployment decision, cost per successful task completion, deteriorated when switching to Chinese across most models tested.
A separate failure mode sits beside that one; multilingual models instructed to reason strictly in Chinese exhibit language mixing during complex logical or mathematical steps, where the model reverts to English unpredictably and produces convoluted outputs that disrupt the very context-saving mechanism the approach was reaching for.
Even where a model’s tokenizer does compress Chinese marginally better than English, agentic workflows do not operate in a vacuum; the agent calls APIs with English parameters, writes code with English syntax, generates artifacts for English-speaking users, and produces logs that English-speaking engineers need to read.
When the agent reasons in Chinese but outputs in English, the model spends tokens translating its internal state at every interface boundary, and that dual-processing step bloats the output context and tends to neutralize whatever savings accrued during the reasoning phase.
The observability cost is harder to quantify, yet it is the one that tends to govern the deployment decision; an agent whose internal reasoning traces are in Chinese is opaque to non-Chinese-speaking developers, and resolving that through real-time log translation requires secondary LLM API calls, which defeats the original economic incentive, so the remaining option is tolerating opaque reasoning in production.
The intermediate translation also introduces generation noise, since each shift between the reasoning language and the output language is a point where semantic precision can degrade; for complex neuro-symbolic tasks where precision is the whole game, the added translation layers raise the risk of hallucination and logic breakdown.
The underlying intuition, that context windows waste capacity on redundant tokens, is correct; the place the fix lives is the tokenizer and compression layers, not the language layer.
Fairer tokenizers. If BPE vocabularies allocated capacity more equitably across languages, the natural semantic density of logographic scripts could translate into real compute savings, and Petrov et al. (2023) frame this as a tokenizer fairness problem.
Algorithmic prompt compression. LLMLingua and similar tools calculate conditional perplexity to identify and drop unnecessary tokens from a prompt; the compression is mathematically grounded, observable, and language-agnostic, and it does not introduce translation overhead because it operates on the prompt’s own language.
Latent-space reasoning. Architectures that process intermediate reasoning steps natively in continuous embeddings or token vectors would sidestep discrete human writing systems altogether.
The hypothesis was reaching for a real inefficiency through the one layer that cannot deliver it; the savings live in the tokenizer, not in the choice of human language the model thinks in.