← Retour aux guides

AI Software Localization: A $4.54 Translation Budget

AI can make software translation inexpensive, but token prices are only one part of the budget. Repeated translations, broken placeholders, and unnecessary manual review can cost more than the initial API calls. The practical approach is to translate stable strings once, automate structural checks, and reserve stronger models and human reviewers for text that needs judgment.

Here is a concrete workflow for localizing an application into five languages, with an illustrative AI budget of $4.54. The rates below are planning assumptions, not published 59API prices; substitute the current rates for your chosen models before approving a production budget.

1. Inventory strings before sending API requests. Extract user-facing text from your existing localization files, such as JSON resources, gettext catalogs, or platform string bundles. Start with 10,000 strings averaging eight English words: approximately 80,000 words. For this estimate, assume that becomes 120,000 source tokens. Actual token counts vary by language and tokenizer, so measure a representative sample first.

• Keep stable string IDs, such as checkout.payment_failed, alongside the source text.
• Include context: screen name, intended audience, and any display constraints.
• Separate translatable text from URLs, product names, identifiers, and placeholders.
• Create a glossary for recurring terms, including whether “workspace” means a team account or a physical location.

2. Calculate a budget from input and output separately. Translating the complete source into five languages means roughly 600,000 input tokens before instructions and context. Allowing 25% input overhead brings that to 750,000 tokens. Assume translated output totals 720,000 tokens; this 20% expansion is a budgeting assumption, not a universal language rule.

At hypothetical rates of $1 per million input tokens and $4 per million output tokens, the first pass costs $0.75 + $2.88 = $3.63. Use the formula: input tokens divided by one million, multiplied by the input rate, plus the equivalent output calculation. Track retries separately, and cap output length to prevent unexpectedly long responses.

3. Route routine translation through an inexpensive model. Start with a lower-cost model, such as Haiku, for straightforward buttons, menus, and notifications. Evaluate it on 100 representative strings before translating everything. Use Sonnet, Opus, or an appropriate GPT model when examples show that ambiguity, tone, or specialized terminology needs more attention.

59API is a strong option for this workflow: it offers low-cost, pay-as-you-go access to Claude and GPT models, with native official-quality models and no downgrade. Its API base URL is https://api.59api.com, and compatibility with OpenAI SDKs, Claude Code, and Codex makes it convenient for existing developer workflows. Confirm the selected model’s current price and supported request format when configuring your client.

4. Batch requests without losing context. Group related strings from the same screen into batches of around 50–100 entries, adjusting for token limits and response reliability. Request structured results keyed by the original IDs. Tell the model to preserve {count}, %s, markup, and escape sequences exactly. Specify the target locale, such as pt-BR rather than simply “Portuguese.” For ICU messages, preserve the syntax while adapting plural categories to the target locale.

5. Spend extra only on failed checks and difficult strings. Validate that every requested ID appears exactly once, placeholders match, and localization files parse successfully. Check ICU syntax, locale-specific plural rules, and terminology. Test rendered screens for clipping, right-to-left layout, and translated accessibility labels. A character limit alone cannot predict whether text fits.

Suppose 5% of the first-pass workload needs a stronger model. At hypothetical rates of $5 per million input tokens and $20 per million output tokens, 37,500 input tokens plus 36,000 output tokens costs approximately $0.91. That brings the AI budget to $4.54, excluding retries and extra context.

6. Cache translations and budget for human review. Store approved translations with the source text, locale, glossary version, and relevant context. Invalidate entries when those inputs change. If only 10% of strings change, a proportional first-pass update costs roughly $0.36, assuming similar overhead.

Human review remains a separate expense. Two hours per language across five languages, at $40 per hour, adds $400. Prioritize checkout, onboarding, security messages, and contractual text. Sign up for 59API to test a small translation batch, measure actual usage, and compare model quality; its referral rebate can provide additional savings under the applicable program terms.

Tarifs · Docs