What AI novel translation really costs
The model charge is one part of a finished translation. A cheap first pass can become expensive when it loses chapters, changes names, or needs a full manual review. Compare workflows with the same source length and required quality.

Estimate model usage, retries, source cleanup, glossary work, review, and ebook production instead of comparing only the price of one request.
Count the source in the unit the service bills
Services can bill by source character, word, token, document, credit, month, or project. Convert the complete source into the same unit before you compare prices. Include titles, notes, and repeated context if the workflow sends them with each chapter.
Token counts vary by language and tokenizer. Do not assume one Chinese or Japanese character equals one token. Use the service’s own estimator when it provides one, then test a real chapter and compare the measured usage.
Add output and repeated context
Some model APIs bill for both input and output. A translation job can also resend the glossary, style rules, character notes, and prior-scene summary with every chapter. That repeated context can become a large part of a long project.
Run a representative chapter from start to finish. Record source tokens, context tokens, output tokens, and any second pass. Multiply by the real chapter count, then add a margin for longer chapters and retries.
Price the work around the model
Source cleanup, OCR correction, glossary preparation, project setup, failed requests, manual review, continuity searches, editing, and EPUB production all take time. If a workflow saves model cost but doubles review time, it can cost more overall.
Give review time an hourly value even when you do the work yourself. This makes a subscription, API workflow, and professional translator easier to compare on the same basis.
- Model input and output
- Repeated glossary and context
- Retries and alternate translations
- Source cleanup and OCR correction
- Meaning review and English editing
- File export, validation, and later corrections
Use a range, not a single total
Calculate a low case with clean source and few retries, a likely case from your chapter test, and a high case with difficult OCR and extra review. A range makes the unknown work visible.
Check the provider’s current pricing and data policy before you start. Model names, rates, limits, and retention terms can change. Do not upload a commercial source unless you have the required rights and the provider’s controls fit the project.


