Send the right unit to the right model.
Translation is not one job. A product attribute is not a legal paragraph. The cascade routes each unit to the model best fit for it, scores the output, and persists.
Six steps. Every unit.
- Step 1
TM lookup
Match against your project's translation memory. ≥0.95 → reuse for free. 0.85–0.95 → use as a hint.
- Step 2
Glossary preprocessing
Brand terms wrapped as model placeholders so they're guaranteed to round-trip unchanged.
- Step 3
Tiered routing
Tier A (DeepL) for short, high-confidence. Tier B (Sonnet) for content. Tier C (Opus) for long-form and structured.
- Step 4
Quality scoring
Claude Haiku rates the output 0.00–1.00. Threshold per project gates auto-approval.
- Step 5
Glossary postprocessing
Placeholders restored to the locale-correct rendering. Brand voice preserved exactly.
- Step 6
Persist
translations row inserted. Embedding indexed for future TM. Webhook fires if status changes.
When each model wins.
| Tier | Default model | Best for | COGS / 1M words |
|---|---|---|---|
| A | DeepL Pro | Buttons, attribute values, menu items, alt text — short and high-confidence | ~£11 |
| B | Claude Sonnet | Post titles, paragraphs, product descriptions, FAQ answers | ~£28 |
| C | Claude Opus | Legal pages, terms, JSON-LD structures, complex marketing copy | ~£72 |
A worked example.
A site translating 1 million words / month, typical mix (60% Tier A, 30% Tier B, 10% Tier C):
- · Weighted COGS: ~£21.50 / month
- · After 60% TM hit rate: ~£8.60 / month
- · Your bill on Growth (£65/mo annual): covers up to 2M words
Numbers are guidance, not guarantees — content mix varies. The point is that the cascade is built around economics, not "send everything to GPT and hope."
Quality always wins where it matters.
The cascade is configurable per project. Route every post_title to Tier C if title quality matters more than spend; route every attribute_value to Tier A even if the engine would have escalated. One YAML.
tiers:
post_title: C # title quality matters more than cost
product_name: C
paragraph: B
attribute_value: A
alt_text: A
json_ld: C # schema must be exact
quality:
auto_approve_threshold: 0.85
flag_for_review_below: 0.70
translation_memory:
reuse_above: 0.95
hint_above: 0.85 Start free in 30 seconds.
No credit card. 100,000 words on the free tier. Self-serve onboarding.