Key facts
| Router model | plugsky-fusion escalates per segment across tiers (live) |
| Translation memory | Exact matches should never be re-translated |
| Glossary checks | Deterministic verification instead of model-based review |
| Models | 30+ models; cheap tiers for high-volume pairs, strong for critical content |
| Streaming | Live via server-sent events where humans watch translations arrive |
| Pricing | Flat monthly self-serve plans with no per-token charges on self-serve |
| Free tier | plugsky-micro and plugsky-lite on the free plan, no card required |
| Roadmap | The batch endpoint is coming soon for large translation runs |
TL;DR
- Exact translation-memory matches should cost nothing.
- Translate only new or changed segments.
- Route by language pair and content type, not one model for all.
- Check glossary and placeholders in code, not with a model.
- Measure cost per approved word, not per API call.
How it works, step by step
- Segment source content with stable ids and index approved translations in memory.
- Apply exact matches automatically and fuzzy matches as suggestions.
- Route remaining segments by language pair and content type.
- Run high-volume UI strings on cheap multilingual tiers with glossaries injected.
- Escalate brand, legal and low-resource pairs to strong models with review.
- Verify placeholders, numbers and glossary terms deterministically before review.
- Feed approved segments back into memory and track cost per approved word.
Try it yourself
Open the LLM token calculator →
Per-segment economics
Translation is priced by volume of segments, and most content is not new. Release notes, product descriptions, help articles and UI strings repeat across versions with small edits. Exact translation-memory matches should be applied automatically at zero model cost, and fuzzy matches should be offered as suggestions that a model refines rather than regenerates.
The remaining work is genuinely new text, and that is where routing applies. High-volume pairs with stable terminology run well on cheap multilingual tiers; low-resource pairs and content where tone matters need stronger models and human review.
Routing by pair and content type
Quality varies by language direction, so a single policy for all translation is either expensive or unreliable. Route by pair and by content type together:
- UI strings and product labels: cheap tier with a strict glossary and placeholder checks.
- Help content and release notes: cheap to mid tier, translation memory assisted.
- Marketing and brand copy: strong tier with human review.
- Legal and regulated text: strong tier, human sign-off, pinned rules.
Store the routing decision with each segment so reviewers and auditors can see which tier produced what.
Checks, memory and measurement
Mechanical errors — a broken placeholder, a translated brand name, a mismatched number — are caught faster and cheaper by code than by a model. Run deterministic checks on every segment and reserve model-based review for tone and fluency questions that code cannot answer.
Feed every approved segment back into translation memory so the next release reuses it. Measure cost per approved word, reuse rate, edit rate after review and time-to-publish. Start on the free plan with plugsky-micro and plugsky-lite for high-volume pairs, then evaluate stronger models per language direction during the 14-day full-access trial; plans are on the live pricing page.
Honest comparison
| Translation choice | Routed pipeline on Plugsky | Strong model for every segment | Human translation only |
|---|---|---|---|
| Exact matches | Translation memory, no cost | Re-translated every time | Reused manually |
| High-volume pairs | Cheap multilingual tiers | Frontier price per segment | Expensive |
| Critical content | Strong tier with review | Native strength | Expert review |
| Mechanical checks | Deterministic code | Model review | Human proofreading |
| Evidence | Cost per approved word | Cost per segment | Cost per word |
Frequently asked questions
How do I reduce translation cost the most?
Reuse first. Exact translation-memory matches cost nothing, and fuzzy matches only need refinement. Translating content that already exists is the largest avoidable expense.
Should all language pairs use the same model?
No. Quality and cost vary by direction. Route high-volume pairs to cheap multilingual tiers after evaluation, and keep brand, legal and low-resource pairs on strong models with review.
Do glossary checks need a model?
No. Placeholder integrity, numbers, do-not-translate terms and glossary compliance are deterministic checks best done in code. Reserve model review for tone and fluency.
How do I handle UI strings?
Use a cheap tier with a strict glossary, preserve placeholders exactly, and cap output length to fit the interface. UI text is repetitive and ideal for reuse.
Is batch translation available?
Not yet — the batch endpoint is coming soon. Run segments through bounded-concurrency workers with a queue and per-segment status today.
How do I evaluate a cheaper tier?
Track edit rate after review, terminology adherence and placeholder integrity per language pair. Move a pair down only when those metrics hold.
How does flat pricing help localization?
Self-serve plans are flat monthly with no per-token charges, so experimenting with tiers and prompts does not create a variable bill. See the live pricing page for plans.
Can I start for free?
Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers stronger multilingual models.