Translation per Yuan: A Two-Model Receipt

01 / The cheaper receipt is not a quality benchmark

For translation per yuan, Gemini 2.5 Flash is the lower-cost starting candidate in this two-model comparison. Claude Haiku 4.5 costs more at the published standard rates below. That establishes a price advantage, not a translation-quality victory: neither a token bill nor this article proves which model translates your language pair better. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing?sid=4323e669-367d-4e2f-b0be-2b51d68ae70f&utm_source=openai))

The useful question is narrower: which model delivers acceptable translations with the least combined spending on generation, retries, and editing? Here is a receipt you can reproduce, followed by the small amount of extra editing that can erase the savings.

02 / The basket: 100 jobs with equal token allowances

Assume 100 translation requests, each using 1,000 input tokens and 1,000 billable output tokens. The basket therefore contains 100,000 input tokens and 100,000 output tokens. Input includes instructions and any glossary. These are budgeting assumptions, not measured translations or claims that equal passages tokenize identically.

For readable yuan arithmetic, assume USD 1 equals CNY 7.20. This is an illustrative conversion, not a current exchange-rate quote. Exclude caching, tools, retries, and extra reasoning from the initial basket; add any actual charges when reconciling your bill.

03 / Receipt one: CNY 2.016 for the Flash basket

Google lists standard Gemini 2.5 Flash text input at USD 0.30 per million tokens and output at USD 2.50 per million tokens, including thinking tokens. Those are vendor reference prices, not verified CheapRelay selling prices. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing?sid=4323e669-367d-4e2f-b0be-2b51d68ae70f&utm_source=openai))

• ■ Input: 100,000 / 1,000,000 × USD 0.30 = USD 0.030.
• ■ Output: 100,000 / 1,000,000 × USD 2.50 = USD 0.250.
• ■ Total: USD 0.280 × 7.20 = CNY 2.016.
• ■ Per translation: CNY 0.02016.
• ■ Per yuan: approximately 49.60 equivalent jobs.

This receipt assumes the full billable output stays within the stated allowance. Do not count only the visible translation if the request also incurs billed thinking output. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing?sid=4323e669-367d-4e2f-b0be-2b51d68ae70f&utm_source=openai))

04 / Receipt two: CNY 4.320 for the Haiku basket

Anthropic lists Claude Haiku 4.5 at USD 1 per million input tokens and USD 5 per million output tokens. Applying those standard vendor prices to the same assumed token basket produces this second receipt. ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing?utm_source=openai))

• ■ Input: 100,000 / 1,000,000 × USD 1 = USD 0.100.
• ■ Output: 100,000 / 1,000,000 × USD 5 = USD 0.500.
• ■ Total: USD 0.600 × 7.20 = CNY 4.320.
• ■ Per translation: CNY 0.04320.
• ■ Per yuan: approximately 23.15 equivalent jobs.

Under these assumptions, Flash costs approximately 53.3% less, saving CNY 2.304 across 100 jobs. That is arithmetic from published prices, not an advertised relay discount. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing?sid=4323e669-367d-4e2f-b0be-2b51d68ae70f&utm_source=openai))

05 / Only 2.76 extra editing seconds erase the saving

Now assume editing labor costs CNY 30 per hour. That equals CNY 0.00833 per second. The generation saving is CNY 0.02304 per job, so approximately 2.76 additional editing seconds per translation would consume it.

This does not mean either model needs more editing. It identifies what to measure. If the cheaper candidate requires noticeably more terminology fixes, its token advantage may disappear. If both need equivalent review, the lower generation receipt wins.

06 / Choose with a blind sample, then reconcile the bill

Test both candidates on the same 100 representative passages. Include product terms, numbers, placeholders, and ambiguous sentences. Hide model names during review. Record accepted translations, correction time, retries, actual token usage, and total yuan charged. Divide total cost by accepted translations rather than submitted requests.

CheapRelay provides one key for both candidates, with pay-as-you-go billing, no monthly fee, and real-token accounting that includes cache reads and writes. Its live model rates remain the missing input here: the vendor receipts above are not a CheapRelay quote. Sign up for CheapRelay to run the sample and replace the reference prices with your actual charges.

My starting choice is Flash for the lower reference-price receipt. My final choice would be whichever candidate produces the lower cost per accepted translation after editing.