Key facts
| Router model | plugsky-fusion escalates per request across tiers (live) |
| Models | 30+ models including multilingual tiers; evaluate Arabic per dialect |
| Embeddings | plugsky-embed-multilingual is live for Arabic and mixed-language retrieval |
| Token effect | Arabic text often uses more tokens than English for the same content |
| Strategies | cost_saver, balanced, max_quality and custom rules per key |
| 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 | Classifier routing (/v1/plugsky/route, model=auto) is coming soon |
TL;DR
- Normalise Arabic text before counting tokens; diacritics and variants add cost.
- Route detection, classification and cleanup to cheap tiers.
- Reserve stronger models for generation, nuance and dialect-sensitive replies.
- Evaluate quality per dialect — MSA results do not transfer automatically.
- Check terminology with a glossary instead of paying a strong model to guess.
How it works, step by step
- Profile real traffic: dialect mix, input length, task types and current model choice.
- Add a normalisation step that standardises diacritics, digits and common variants.
- Route detection, language ID, classification and formatting to plugsky-micro or plugsky-lite.
- Escalate generation, summarisation and nuance-heavy tasks to stronger multilingual tiers.
- Use plugsky-embed-multilingual for retrieval and deduplication across Arabic and mixed content.
- Verify terminology with a glossary check rather than relying on the model alone.
- Build dialect-specific evaluation sets and re-measure after every routing change.
Try it yourself
Open the AI model pricing calculator →
Why Arabic requests cost more
Arabic is morphologically rich: a single word can carry what English expresses with several, and the same sentence can be written with or without diacritics, with different digit sets, or with dialect markers that change meaning. Tokenisation reflects that complexity, so the same semantic content often consumes more tokens than English, and small models make more mistakes on dialect-heavy text.
The cost response is not to avoid Arabic models — it is to normalise and tier. A normalisation pass that standardises diacritics, Arabic-Indic digits and common variants shrinks inputs and improves consistency. Detection, classification and cleanup then run cheaply, while the expensive calls concentrate on generation and reasoning.
Routing by dialect and task
Modern Standard Arabic, Gulf, Egyptian and Levantine content behave differently. Build routing around both task and dialect rather than treating Arabic as one language:
- Cheap tier: language and dialect detection, sentiment, routing, short replies, formatting.
- Mid tier: summarisation, translation between Arabic variants, first-draft generation.
- Strong tier: nuanced writing, legal and religiously sensitive content, final customer-facing copies.
- Ensemble or review: high-stakes outputs where a second pass is cheaper than a bad answer.
Fusion strategies set the default per key, and custom rules pin paths that must not be downgraded. Classifier routing with model=auto is coming soon.
Evaluation and retrieval in Arabic
Quality must be measured per dialect. A model that performs well on MSA can struggle with Gulf colloquial, and evaluation sets translated from English miss dialect phenomena entirely. Collect real examples, label expected outputs, and score terminology, tone and correctness per variant.
For retrieval, plugsky-embed-multilingual handles Arabic and mixed-language corpora, which matters in GCC enterprises where documents, queries and product names mix scripts. Index Arabic content with normalisation applied and keep a glossary check in the pipeline so brand and regulatory terms stay exact. Start on the free plan with plugsky-micro and plugsky-lite, then use the 14-day full-access trial to compare stronger tiers on your dialect sets; the live pricing page lists current plans.
Honest comparison
| Cost lever | Arabic routing on Plugsky | One strong model for all | One cheap model for all |
|---|---|---|---|
| Detection and routing | Cheap tiers after normalisation | Strong tier does trivia | Cheap tier makes errors |
| Dialect nuance | Escalated to stronger tiers | Handled natively | Often wrong |
| Normalisation | Removes variant overhead early | Still pays for noise | Still pays for noise |
| Retrieval | plugsky-embed-multilingual | Depends on provider | Depends on provider |
| Evaluation | Dialect-specific sets | Usually MSA only | Usually MSA only |
Frequently asked questions
Why does Arabic AI cost more than English?
Arabic morphology and orthographic variation often produce more tokens per unit of meaning, and dialect text is harder for small models. Normalisation and tiered routing address both effects.
Should I always use the strongest model for Arabic?
No. Use cheap tiers for detection, classification and cleanup, and reserve stronger models for generation and nuance. Evaluate per dialect before deciding where the boundary sits.
Which embedding model handles Arabic?
plugsky-embed-multilingual is built for multilingual and mixed-script corpora, which matches GCC document collections where Arabic, English and product names appear together.
How do I handle multiple dialects?
Detect the dialect, route accordingly, and evaluate on real examples per variant. Do not assume Modern Standard Arabic performance transfers to Gulf or Egyptian content.
Does normalisation change meaning?
A conservative pass that standardises diacritics, digits and common letter variants preserves meaning while reducing noise. Keep the original text for audit and apply normalisation only to model inputs.
Is classifier routing available?
Not yet — model=auto classifier routing is coming soon. Today, fusion strategies and custom rules per key cover most Arabic routing needs.
How do I keep terminology consistent?
Maintain a bilingual glossary, check outputs against it deterministically, and retrieve approved terms during generation. Glossary checks are cheaper than a stronger model guessing.
Can I evaluate on a free plan?
Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial lets you compare stronger multilingual tiers.