What's new
Explainer14 Nov 20265 min read

Budget translated prompts from their token counts

Two prompt versions with different token counts cost different amounts to send, even at the same price and request volume, so count the exact request for your model instead of borrowing a budget from another language.

Say you localise an assistant's prompt and keep the monthly budget. The translated wording looks about as long as the original, so the old estimate stays in the spreadsheet. A count of the actual request is better than that guess. It should cover the system instructions, the examples and any returned material, as well as the latest question.

Anthropic's token-counting guide, read on 14 November 2026, describes a way to estimate structured message input for a chosen model. It works for the supported input forms, and a real trial will show how close the estimate runs to reported usage. We do not claim that an estimate is an exact future bill.

Why a language label is not a ratio

The table uses versions A, B and C with assumed lengths of 4,000, 7,000 and 12,000 tokens. They are not measurements for English or any other language. At the same invented input rate and monthly volume, the bills differ only because the stated token counts differ.

Synthetic prompt lengths at an assumed $1 per million input tokens. No language-specific tokenisation or translation quality was measured.
Fictional prompt versionAssumed input tokens/requestMonthly requestsInput-only bill, USD
A400010000$40.00
B700010000$70.00
C1200010000$120.00

Ten thousand requests at 4,000 input tokens come to forty million input tokens. At 12,000 they come to 120 million. The $40 and $120 bills are input-only illustrations and say nothing general about the ratio between languages or models.

What goes into a representative count

The complete localised prompt, examples included, needs its own version label. Count it with realistic user input and documents, in the form the application will send. Translating the system instructions alone does not represent the complete request.

The count belongs with the target model identifier. A count made for one model should not move into another model's budget without checking how that model tokenises. For input types the counter does not support, the provider's documented method or an observed usage record is a better source than a plain-text count.

One size is rarely enough. A short question and a long attached document stress different parts of the request, so the range seen in the trial is more honest than a single figure presented as the prompt size.

Did the task survive translation?

Each language version needs reviewed acceptance examples, meaning sample inputs with answers someone has agreed are correct. They should keep the intended constraints, such as which fields to extract and how to express missing information. A shorter prompt that drops a condition is a different task, so its lower token count is not a like-for-like saving.

Local date, number and address formats matter where the application depends on them. Someone, or some process, qualified in the language and task should review those outcomes. We propose that check but have not evaluated translation quality here.

Output is a separate measurement. A translated instruction can produce a differently sized answer, which an input-only estimate cannot capture. Accepted answers and their actual output usage need counting before the full monthly workload is forecast.

Closing the gap between estimate and usage

A small authorised trial, with the request version kept, gives a usage record to read. Any gap between the preflight estimate and the observed input should be explained before the budget scales. Tool definitions or other content added later go into the next estimate.

For public evidence, synthetic examples and aggregate counts are enough. Real customer messages or private translations do not belong in a repository just to make the comparison reproducible.

A row in your own results log needs the localised version, model, counting method, request mix and date checked. That lets later costs be compared under clear assumptions, with language quality kept as a separate result that still needs review.