AI-Assisted i18n Translation Cost: Token Math per 1,000 UI Strings (Claude, GPT, Gemini 2026)

Why "It's Just Text" Is Wrong

The naive calculation for translating 1,000 UI strings looks like this: average English string is 8 words ≈ 12 tokens. Add the same for the translated output. Times 1,000. That is 24,000 tokens, or 24 cents on Claude Sonnet. It is a very wrong number.

The real cost multipliers, in order of impact:

1. Context per call. You cannot translate an ICU MessageFormat string safely without passing the format specification, the placeholder semantics, and often the containing screen context. That is 300-500 tokens of overhead per call.
2. Brand terms and glossary. The product name, feature names, and industry jargon must appear verbatim. A glossary of 50 terms plus rules for when to translate them adds 400-800 tokens per call.
3. Verification. Any production i18n pipeline runs at least one verification pass to catch placeholder misalignment, character-limit violations, and RTL issues. That is another full round-trip.
4. Ambiguity resolution. "Save" could mean rescue or persist-to-disk. Well-designed UI systems pass screen/tooltip context; that is another 100-200 tokens per string.

Realistic Per-String Token Count

Bundling in the multipliers, a production-quality LLM translation of one string typically consumes:

• Input: ~1,200 tokens (system prompt + glossary + context + string).
• Output: ~30 tokens (the translation).
• Verification pass: ~800 input + ~30 output tokens.
• Per string total: ~2,060 tokens.

Batching (10 strings per call, shared context) reduces the per-string overhead significantly. With batching, per-string effective cost drops to ~350-500 tokens. The math below assumes 400 tokens per string with batching enabled.

1,000 UI Strings, Per Target Language, By Model

For a product supporting 10 languages, the total translation bill runs $6-$120 depending on model choice — often cheaper than one hour of a professional translator's time.

Where the Hidden Cost Bites

Three failure modes generate expensive rework:

• Placeholder drift. The model rewrites {userName} as {utilisateur} in French. UI breaks at runtime. Verification pass or strict schema output catches this.
• Length overflow. German translations are on average 30% longer than English source. Buttons and menu items overflow. Pass a max-character constraint per string.
• Plural forms. Slavic languages have 3-4 plural forms; Arabic has 6. Naïve string-per-string translation loses this. Use structured plural output (ICU MessageFormat).

Each failure caught in verification instead of in production saves ~$40-$80 of engineering time. That is why the verification pass — which doubles LLM cost — pays for itself many times over.

Recommended Setup

1. Store source strings in an ICU MessageFormat-compatible catalog (JSON, XLIFF, or Fluent).
2. Batch 10-20 strings per LLM call, sharing the system prompt and glossary.
3. Use Sonnet 5 or Gemini 3 Pro for the translation pass, Haiku for a verification pass.
4. Add a schema-enforcing wrapper so the model cannot break placeholder syntax.
5. Cap per-string retries at 3; escalate failures to a human review queue.
6. Cache translations by source-string hash + glossary version so re-runs are free.

When to Skip LLM Translation Entirely

LLMs are cheaper than professional translators but not always the right tool. Skip LLM translation when:

• Regulatory content requires certified translation (medical, legal, finance).
• Marketing hero copy where brand voice matters more than throughput.
• Languages where the model's training data is thin (some African, Indigenous, low-resource languages).

Everything else — settings screens, error messages, form labels, tooltips — LLM translation at $1-$3 per 1,000 strings is the current price floor and it is unusually good value.