A translation quote per thousand words sounds straightforward until the API bills in tokens. Source text, translated output, instructions, and revision requests all contribute to the bill. This first-run checklist turns those moving parts into a repeatable estimate, with human review included rather than hidden outside the budget.
1. Define one thousand source words as your unit
Start with 1,000 source words, not 1,000 translated words. Target-language length can change, so using the source count keeps estimates consistent across projects.
Write down the language pair, content type, and required deliverable. A product description needing light editing has a different workflow from a legal document requiring specialist review. Also decide whether headings, repeated passages, and glossary entries count toward the source total.
2. Measure tokens on a representative sample
Do not treat words and tokens as interchangeable. Tokenization varies by language, punctuation, formatting, and model. Select a representative sample, translate it with your intended settings, and inspect the usage returned by the API.
For this walkthrough, suppose a 1,000-word sample produces these illustrative counts:
- Source text: 1,400 input tokens.
- Instructions and glossary: 200 input tokens.
- Translation: 1,700 output tokens.
That gives 1,600 input tokens and 1,700 output tokens. These are planning assumptions, not universal conversion ratios. Replace them with measured usage before quoting a real job.
3. Apply separate input and output prices
Record the current prices for the exact model and provider route you intend to use. Check the billing unit carefully: prices quoted per million tokens must be divided by 1,000,000 when calculating a request.
Translation charge = (input tokens × input price per million + output tokens × output price per million) ÷ 1,000,000.
Using hypothetical rates of $0.50 per million input tokens and $2.00 per million output tokens, the input costs $0.0008 and the output costs $0.0034. The total is $0.0042 per thousand source words. These rates illustrate the calculation; they are not a SwiftRelay price quote.
4. Count revision requests and repeated instructions
A first draft is not always the final deliverable. If your workflow includes a terminology check, back translation, or rewritten paragraphs, include those requests in the estimate. A revision can bill the original text, the draft, and fresh instructions as input.
Measure a complete workflow rather than multiplying the first request by a guessed factor. Also watch chunking: splitting a document into many requests may repeat the same glossary and instructions. That repeated context increases the effective cost per thousand words.
5. Add review time to the translation budget
API charges are only one part of a usable translation estimate. Track the time needed to check meaning, terminology, formatting, and omissions. For sensitive material, budget for an appropriately qualified reviewer.
Suppose review takes three minutes per thousand source words at $30 per hour. Review then costs $1.50. Added to the illustrative API charge of $0.0042, the working total becomes $1.5042. A 10% contingency brings it to $1.65462, or about $1.65 per thousand source words. Keep extra precision internally and round only the final estimate.
6. Validate the workflow through SwiftRelay
SwiftRelay is a practical option for testing this calculation with pay-as-you-go, per-token billing. It supports OpenAI-compatible SDKs, so an existing translation script can use https://api.fastapirelay.com/v1 as its base URL. Choose a candidate such as Gemini 2.5 Flash, then assess translation quality and actual charges on your sample.
Sign up for SwiftRelay to use the small gift credit for an initial translation trial. Treat that credit as a testing allowance, not a reduction in your ongoing unit-cost estimate.
7. Save a repeatable thousand-word estimate
Keep a record of the source word count, language pair, model, pricing date, input tokens, output tokens, revision charges, and review minutes. For a larger completed test, divide the total cost by source words and multiply by 1,000.
Your final quote should state what it includes. A defensible estimate is not just a token calculation: it is a measured cost for delivering reviewed translation at a defined quality level.