「英語で使えば安い」は本当か？ Claude Opus 4.8 のトークン消費を3言語で比較検証 An experiment measuring Claude Opus 4.8 token consumption across English, Japanese, and Ch…

The idea that using English with AI models saves money is a common piece of developer wisdom, but hard numbers are surprisingly rare. This experiment puts the claim to the test by measuring output token consumption from Claude Opus 4.8 across three languages — English, Japanese, and Chinese — while holding the informational content of the generated summaries constant.

The results largely confirm the intuition. Taking English output as the baseline, Japanese required approximately 1.23× the token count, and Chinese came in at roughly 1.29×. In practical terms, that means generating the same content in English could cut costs by around 20% compared to Japanese and nearly 30% compared to Chinese — a meaningful difference when operating at scale with a flagship model like Opus 4.8, which carries a premium per-token price.

The root cause lies in how tokenizers work. Most large language models, including Claude, use subword segmentation algorithms such as Byte Pair Encoding (BPE), which are optimized around English vocabulary patterns. English words tend to map to single or compound tokens efficiently, while Japanese and Chinese characters are often split at a much finer granularity — sometimes individual characters or morphemes — requiring more tokens to encode the same semantic content. This is not a quirk unique to Anthropic; OpenAI's tiktoken and most other tokenizers share the same structural bias toward alphabetic scripts.

One nuance worth noting is that input tokens are subject to the same dynamic. Writing prompts in English could also reduce input costs, compounding the savings. However, the practical calculus isn't purely about token counts. Translating prompts and outputs introduces friction, potential quality degradation, and engineering overhead. For tasks requiring nuanced understanding of Japanese cultural context or domain-specific Japanese terminology, forcing an English-language pipeline may introduce errors that outweigh any cost benefit.

There's also the broader optimization picture to consider. Developers serious about cost control typically have more leverage in model selection than in language choice. Switching from Opus 4.8 to a smaller, cheaper model in the same family likely yields greater savings than switching languages alone. The most cost-effective strategy likely combines both levers — using a lighter model when quality permits, and defaulting to English output where quality is not compromised.

It's worth flagging that the source experiment appears to be based on a limited sample set, and the observed ratios could shift depending on task type, text length, and domain. Technical documentation, for instance, may show different ratios than conversational summaries. That said, the directional finding — English is the most token-efficient output language for current LLMs — is structurally well-supported by the underlying tokenizer design and unlikely to reverse absent a major architectural shift.

For teams running high-volume Claude workloads, language selection deserves a seat at the cost-optimization table alongside model choice, prompt compression, and caching strategies. This experiment provides a concrete, if preliminary, data point for those conversations.