Start with a representative page. AI usage can be a small part of a multilingual project, but page count alone does not predict the bill. Measure the input and output tokens for your content, multiply by your target languages, and budget separately for review and the Omni licence.
What actually costs money?
The Omni Single Site Licence costs £249 per year. AI usage is billed separately by the provider you connect through Umbraco.AI. Content preparation, editorial review and publishing are also part of your project budget.
Providers commonly price input and output tokens separately. Input includes the source text and instructions; output includes the generated translation. Token counts vary by language and model. Check your provider's rules for reasoning tokens, caching, taxes and any additional charges.
Use the official Claude pricing documentation or Gemini API pricing for the exact model and service you intend to use. A rate from one provider or hosting platform does not automatically apply to another.
How Omni sends content to the model
Omni extracts translatable text units from supported properties. A unit might be a heading, a rich-text paragraph or a text value inside a block. Requests include translation instructions as well as that text. The package preserves the surrounding content structure when applying the results.
That repeated instruction overhead matters, especially for short strings. Its size depends on the configured prompt, glossary, content and model tokenizer. Retries and rich-text fallback work can add further requests.
Short strings have proportionally more overhead
For illustration, assume 300 instruction tokens per request. These figures show the share of input used by instructions; they are not measured defaults for every installation.
| Example text unit | Source tokens | Instruction share of input |
|---|---|---|
| Dictionary label | 8 | 97% |
| Heading | 20 | 94% |
| Paragraph | 60 | 83% |
| Long rich-text section | 400 | 43% |
Short labels can also be ambiguous without context. Low token usage does not remove the need to check terminology and meaning.
A worked page estimate
Assume 25 text units, each with 60 source tokens and 300 instruction tokens. Assume the translated output also averages 60 tokens per unit, with no retries, caching or extra reasoning charges.
- Input: 25 × (60 + 300) = 9,000 tokens.
- Output: 25 × 60 = 1,500 tokens.
Cost = (input tokens × input rate + output tokens × output rate) / 1,000,000The following are example rate pairs in US dollars, not a current model ranking. Substitute your provider's rates. Keeping the calculation in the billing currency avoids hiding an exchange-rate assumption.
| Example input / output rate per million tokens | One page, one language | 200 pages × 3 languages | 500 pages × 5 languages |
|---|---|---|---|
| $0.10 / $0.40 | $0.0015 | $0.90 | $3.75 |
| $1 / $5 | $0.0165 | $9.90 | $41.25 |
| $2 / $10 | $0.0330 | $19.80 | $82.50 |
The larger totals assume every page has the same text volume as the example. A long page with many blocks, more verbose output or repeated retries can cost considerably more.
Dictionary example
For 500 missing Dictionary values translated into three languages, assume 8 source tokens, 300 instruction tokens and 8 output tokens per value. That is 462,000 input tokens and 12,000 output tokens. At the example $1 input / $5 output rate, the total is $0.522 before any additional charges.
Dictionary values are saved directly, without a draft publishing stage. Start with a small selection to check the wording and workflow, not simply because the usage cost is low.
Choose a model by the result, not just the rate
Compare candidate profiles on the same sample of your real content. Include short labels, brand terminology, long rich text, blocks and the language pairs you need. Have a fluent reviewer assess meaning and editing effort alongside latency and usage cost.
| Content | What to evaluate |
|---|---|
| Dictionary and navigation | Ambiguity, consistency and length constraints |
| Informational pages | Accuracy, completeness and readable phrasing |
| Marketing copy | Brand voice, idiom and cultural fit |
| Specialist or regulated material | Suitability and review by someone qualified in the subject; model price is no guarantee of accuracy |
| Each source / target language pair | Terminology, regional usage and the rendered site's layout |
Omni supports multiple approved Umbraco.AI profiles, so you can use different models where testing supports that choice. The provider selection guide covers the wider evaluation.
Estimate your own site
- Choose representative saved pages, including your longest and most complex content.
- Translate a small sample into each required language and record provider-reported usage.
- Include retries and failed requests in the measured cost, and record the editorial time separately.
- Scale by the content types and page counts you actually intend to translate.
- Add an explicit allowance for reruns and maintenance, then check usage as the rollout progresses.
Use provider-side budget controls or alerts where available, and check whether the control stops requests or only sends a notification. See provider configuration before running a large section job.
Common questions
Is AI usage included in the licence?
No. The annual Omni licence covers the package. Provider usage is a separate bill.
Do updates incur more AI charges?
Yes, when you run another translation job. Source-change tracking helps target the work, but does not make requests free.
Can I use free provider credit?
Check the provider's current eligibility, limits and data terms. Do not budget a production rollout on the assumption that trial credit will remain available.
Plan the wider project
Review Omni pricing and requirements and planning translation as an agency service to account for configuration, editorial review and ongoing maintenance.