Genshijin coding assistant trimmed Japanese code responses by roughly 28 % but nudged English replies 2.5 % longer, and in both cases the total API bill went up.

Anyone paying for AI-generated code feels the impact. Token-based pricing turns every extra word in a prompt or output into dollars spent. A tool that shortens one language while inflating another can swing cost calculations dramatically.

How the test was run

The experiment covered 192 software-engineering prompts in Japanese and English. Researchers sent each query to the model under four reply configurations: a plain baseline and three variations that invoked the Genshijin skill with different instruction sets. The skill works by appending extra guidance to the prompt, which consumes additional input tokens.

They measured output length in tokens and added the cost of input and output tokens to get the total API cost.

What the numbers say

  • Japanese output fell by 27.5 % when Genshijin was active.
  • English output grew by 2.5 % under the same conditions.
  • For both languages, total API spend rose.

The cost increase comes from the extra instructions Genshijin injects. Those instructions add tokens before the model even starts generating code. In this test the token savings from shorter Japanese output never outweighed the extra input tokens required.

To break even, a Japanese query would need to shave off about 479 output tokens, while an English query would need a 435-token reduction. The measured savings fell short of those thresholds.

Why the language gap?

Japanese often carries padding in the form of honorifics and verbose phrasing. Genshijin’s prompt tweaks target that padding, stripping unnecessary words and compressing the response. English, by contrast, tends to be more concise already, so the same prompts can inadvertently add boilerplate that lengthens the answer.

What developers should keep in mind

  • Long Japanese explanations: If you routinely request detailed design notes or extensive comments in Japanese, Genshijin can trim fluff and lower the token count of the output.
  • Short queries: For brief questions, a simple “answer concisely” instruction beats the full Genshijin skill, yielding shorter replies and lower costs.
  • English-heavy workflows: Expect longer outputs. Measure the impact before assuming a cost win.
  • Measure yourself: Token usage varies with prompt style, model version, and task complexity. Run your own benchmarks rather than relying on advertised percentages.

The 75 % reduction some documentation touts refers only to the output token count, not the total bill. Developers who focus solely on that figure may be surprised when their invoices stay the same or even climb.

What to watch next

Takeaway: Genshijin can meaningfully shrink Japanese code output, but the extra prompt tokens it adds erase any immediate cost benefit—and it can even inflate English responses. Use it selectively, and always verify token impact on your own workload.

Source: https://dev.to/anicca_301094325e/coding-assistant-tests-japanese-28-shorter-english-3-longer-4f83