¥1 Translation Budget: Rank by Usable Output

A cheap translation stops being cheap when you pay to fix it twice. If your budget is ¥1, the useful question is not which public model sounds most impressive. It is which model translates the most usable text for the money actually deducted from your balance.

You can build that ranking yourself with a small, controlled test. The spending amounts below are planning limits and hypothetical receipts, not current model prices or published benchmarks.

01 / Split your ¥1 before sending any text

Give the first comparison a ceiling of ¥0.20. Keep ¥0.60 for useful work with the winner and ¥0.20 for corrections. These are spending limits, not promises about how much translation each amount buys. Check the displayed rates before making requests.

Start with two candidates, such as Gemini 2.5 Flash and Claude Haiku 4.5. Their inclusion here is a shortlist, not a claim that either is cheapest or best. Testing two keeps the experiment manageable; a ten-model tournament can consume your budget before you translate anything useful.

CheapRelay provides pay-as-you-go access through one API key, with no monthly fee or subscription. That makes it a practical place to run a small comparison, while keeping the decision tied to the actual charges for your requests.

02 / Use 1,000 source characters from your real work

Choose five passages of roughly 200 source characters each. Include ordinary prose, a sentence with ambiguous meaning, numbers or dates, specialist terminology, and any formatting you need preserved. Use non-sensitive material you can legally submit.

Keep the translation direction fixed. A ranking for English into Chinese does not establish the winner for Chinese into English. Use identical passages and instructions for both candidates.

A plain prompt is enough: Translate the following text into English. Preserve names, numbers, units, paragraph breaks, and meaning. Return only the translation. Add your preferred terminology if you already have it.

03 / Count accepted source text, not impressive output length

Define acceptance before reading the results. For this beginner test, accept a passage only when it preserves the meaning, contains every important detail, and needs no substantive correction. A prettier sentence should not outweigh a missing negation or an incorrect price.

Compare each translation with the source. If you cannot judge both languages, ask someone who can to review the small sample. Fluency alone cannot establish accuracy. Any paid review also belongs in a separate total-cost calculation.

Use accepted source characters as the output measure. Counting translated characters can reward wordiness and distort comparisons between languages. Your score is accepted source characters divided by total yuan charged, including paid retries.

04 / A ¥0.06 receipt can beat a ¥0.09 receipt

Here is an invented example to show the arithmetic. It does not represent measured performance by either named model.

• ■ Candidate A costs ¥0.06 across its initial requests and corrections. Four passages pass, representing 800 source characters.
• ■ Candidate B costs ¥0.09 across its initial requests and corrections. All five pass, representing 1,000 source characters.
• ■ Candidate A scores about 13,333 accepted source characters per yuan. Candidate B scores about 11,111.

On spending efficiency, A ranks first. On complete coverage of this sample, B ranks first. If every passage must pass, B is the eligible choice. Set that requirement before choosing a winner.

05 / Read the token bill before scaling up

Character counts measure useful work here; they do not predict the invoice. API billing uses tokens, and your instructions, source text, response, and applicable cache activity affect the charge. CheapRelay bills real tokens, including cache reads and writes.

Record the model identifier, displayed rates, actual charge, accepted characters, and correction attempts for each test. Avoid assuming that the same yuan amount always buys the same number of words.

06 / Spend the remaining ¥0.60 on the measured winner

Sign up for CheapRelay, check the current rates, and run the small comparison within your chosen ceiling. Then send useful work in batches, inspect the first results, and stop if quality slips. Your ranking should describe a specific language pair and task, backed by receipts you can explain.