Classic machine translation gives you a sentence-for-sentence rendering and no say in how. It doesn't know that your brand voice is warm but never jokey in German, that "Premium Plan" stays in English in every market, or that your Japanese support emails use a specific level of formality. Claude approaches translation as an instruction-following task: you state those rules in the prompt, and they apply to every sentence. For enterprises, that shifts translation from "words per dollar" to something closer to what human localization teams actually deliver — at a scale and turnaround human teams can't match, with humans moved to the review seat.
Terminology control: the glossary goes in the prompt
The core technique is disarmingly simple. Your system prompt carries the standing rules: the glossary (terms that must translate a fixed way, and terms that must not be translated at all), tone guidance per market, formality register, formatting conventions for dates and currency, and a few example passages translated the approved way. The content to translate goes in the user message. Because the glossary rides along with every request, consistency stops depending on which translator — human or machine — happened to pick up the file.
That system prompt gets long: a real glossary plus style guide can run to thousands of tokens, repeated on every request. This is precisely what prompt caching is for. Available on all four platforms — Bedrock, Vertex AI, Foundry, and Claude Platform on AWS — caching lets the platform reuse the long, unchanging prefix across requests so you are not paying full price to resend your style guide ten thousand times a day. For catalog-scale jobs, it is routinely the difference between a rounding error and a budget line.
Tiering: not all content deserves Opus
Translation workloads are wildly uneven in stakes, and your model choice should follow suit. Bulk, low-risk content — product listings, internal documentation, support-ticket gists — runs well on Claude Haiku 4.5, the fast, low-cost tier. Customer-facing prose where nuance matters — marketing copy, executive communications, sensitive support replies — justifies Claude Sonnet 5 or Opus 4.8. A simple router (content type in, model choice out) captures most of the savings; see model tiering for the general pattern. One operational note: if you planned to run overnight bulk jobs through the Batch API, it is not available on Bedrock or Vertex AI — on those platforms, batch-style work means orchestrating ordinary requests yourself.
Localization is more than translation — say so in the prompt
Localization means adapting, not just converting: units, date formats, cultural references, idioms, regulatory phrasing. Claude handles this well when asked explicitly. A prompt that says "adapt idioms and examples for a French business audience; convert measurements to metric; flag any passage whose meaning you had to change to localize it" produces materially better output than "translate to French." That last clause — asking the model to flag its own adaptations — gives reviewers a shortlist of exactly the places human judgment is needed.
Pitfalls seen in real rollouts
No source of truth for terminology. If three departments keep three glossaries, the model will be consistently inconsistent. Consolidate the glossary first; it pays off across human and machine translation alike.
Measuring nothing. Build a small evaluation set — a few dozen representative passages with approved translations — and score new prompts and model versions against it before rolling them out. Otherwise "the German feels off" debates never end.
Translating strings without context. UI strings like "Back" or "Save" are ambiguous in isolation. Send surrounding context (screen name, neighboring strings) with each string; quality jumps immediately.
Forgetting round-trip checks. For high-stakes content, translating back to the source language and comparing is a cheap automated sanity check that catches meaning drift before a human ever looks.
Letting formatting rot. Markup, placeholders like {name}, and variable tokens must survive translation intact. Validate them in code after every call — this is a deterministic check, not a judgment call.
Where to go next
Prompt caching is the cost lever that makes glossary-heavy translation affordable — see prompt caching for how it works, and model tiering for routing content to the right tier. The model lineup summarizes the three tiers at a glance.