AI settings and usage

plinth can use AI to write product texts, suggest categories, describe images and more. For that the company enters its own API key. In the AI settings you decide which AI features run, with which model and up to which monthly budget. The page AI usage shows what the AI has cost.

Two rules always apply: only product data is sent to the AI, and AI results are always suggestions that a person reviews, never direct changes.

Where to find it

• Settings: Help and administration > AI settings.
• Usage: Reports > AI usage.
• Both pages share the tabs Usage and Settings.
• The key can also be entered in the wizard, step AI. See Onboarding wizard.

What you need

The AI settings apply to all company shops. To change them you need Manage AI in every shop of the company. Otherwise you see This setting applies to all shops of the company. It needs the permission in every shop of the company.

You also need an API key from Anthropic. The company orders and pays for it itself.

How do I enter the AI key?

1. Open Help and administration > AI settings.
2. In the section API key, paste the key into Add key. It starts with sk-ant-.
3. Click Save. The key is stored encrypted and never shown again.
4. Click Test key. plinth sends a small test request with the stored key.
5. The line Status now shows Works with the date of the test.

Without a key the status shows Rules only: the AI features are off and plinth works with rules only.

How do I replace or remove the key?

1. To replace it, paste the new key into Replace key and click Save. A new key replaces the old one immediately. Test it again.
2. To remove it, tick Yes, remove and click Remove key. All AI features are off afterwards.

How do I switch AI features on or off?

1. Open Help and administration > AI settings and go to the section Features.
2. Each line is one AI feature with a short description. Use the switch in the column On. Off means: this feature sends nothing to the AI.
3. Choose the Model for the feature: Haiku, Sonnet, Opus or Fable. Larger models are more accurate but cost more. Haiku is enough for checks and short texts. Below the choice you see the version that runs right now, for example currently Sonnet 5.5. You can leave the Effort on Automatic. See AI models.
4. Click Save. You see Features saved.

A feature marked (default) has not been set by you yet; the default applies. Which features are listed depends on the areas that are active for the shop.

Note: The instructions the AI follows when writing product texts (text style and prompts) are not set here. You find them in the product settings. See Products.

AI models: which model for which task, and what it costs

plinth works with four models from Anthropic. You only choose the model, never a version number: plinth always uses the newest version of the chosen model automatically, as soon as it is available for your key. The version that runs right now is shown in small print next to the choice, for example currently Sonnet 5.5.

Recommendation per task

All amounts are estimates as of 12 November 2026: Anthropic list prices, a rate of €0.86 per US dollar, typical amounts of text per product, effort Low. Your actual costs depend on the length of your product data and texts. What was really spent is on the page AI usage.

For comparison, what the same task costs with another model, using 1,000 product texts as the example (estimate as of 12 November 2026): Haiku about €0.50, Sonnet about €10, Opus about €20, Fable about €41.

How the numbers come about: one product text needs about 2,400 input tokens (instructions, text style, product data) and 700 output tokens. With Sonnet that is 2,400 × $2.00 + 700 × $10.00 per 1 million tokens = $0.0118, about 1.0 cent.

What lowers the cost:

• Cache: the same request is never paid twice. Unchanged products cost nothing on the next run.
• Batch: runs that are not urgent cost half as a batch. plinth shows the batch price in the estimate where a batch is possible.
• The right model: Haiku for assignments and checks, Sonnet for texts. Opus and Fable only where the result with Sonnet is not good enough.

Cutting out images does not use any of these models and is not in this table. See Images.

What does effort mean?

The effort sets how thoroughly the model thinks before it answers.

Higher effort costs more because the thinking is charged as output tokens. The amounts above are for Low. Only raise the effort when the results are not good enough.

Where do I choose the model?

• As the company setting: under AI settings > Features, per feature in the columns Model and Effort. Automatic for the effort means: the default of the model.
• Right where you start: wherever you start AI work, plinth shows the same choice of Model and Effort, for example for the AI texts of an import, the AI run of the products, Rework existing products, the image check, shop the look, the category suggestions of data quality and the brand research. The company setting is preselected. For runs with a cost confirmation, plinth shows the new price after another choice before you confirm. The choice applies to this one run only.

Which version is running right now?

The section Models in the AI settings shows the current version of each model, what it is meant for and the list price.

plinth reads the model list again in the background as soon as it is older than one day, and on every Test key. Refresh now reads it right away. This costs nothing. If the provider cannot be reached, the last known list still applies.

If a certain version must stay, open Set an exact model ID (advanced) and enter the ID, for example for Fable. Clear the field and plinth uses the newest version again.

How do I see the real costs?

1. Open Reports > AI usage.
2. By feature shows what each task cost in the month, By model what each model version cost.
3. If a model is marked price estimated, the price list does not know this new version yet. plinth then calculated with the price of the last known version of the same model.
4. With a monthly budget and Hard stop you limit the costs. See above.

How do I set a budget?

1. Open Help and administration > AI settings and go to the section Budget.
2. Enter the Monthly budget (EUR). 0 means no budget.
3. Enter Warning levels (%), for example 50, 80. When a level is reached, the usage page shows it.
4. Decide on Hard stop: when switched on, new requests are rejected once the budget is used up. Results from the cache stay free.
5. Click Save. You see Budget saved.

The budget is a monthly company budget. Costs are estimates based on the price list, not an invoice.

How do I see what the AI costs?

1. Open Reports > AI usage.
2. Choose the Month and the Scope (All shops of the company or This shop only) and click Show.
3. The four tiles show Cost, Requests, From cache and Tokens for the month.
4. Budget shows how much of the monthly budget is used. Change budget opens the settings.
5. The tables below show the usage By feature, By model, By month (the last 12 months), Most expensive products and By day.
6. With View AI usage per person you also see By person. This is meant to find problems, for example a process that starts too often, not to rate people.

How do I estimate the cost before a large run?

1. Open Reports > AI usage and expand Estimate cost.
2. Choose the Feature and enter the Number of products. With Model and Effort you can work out what another model would cost; Company setting takes what is set for the feature. Adjust the other fields if you know better values.
3. Click Estimate. Nothing is sent to the AI and nothing is charged.
4. The result shows the approximate number of requests, tokens and cost. It is an upper limit; the actual output is usually much shorter.

What is the reminder "Review AI usage"?

From time to time the usage page shows Review AI usage: take a quick look at costs and features. Is everything worth it, or can anything run cheaper? Click Reviewed when you have looked. The reminder returns in 30 days.

Which data goes to the AI?

• Only product data: names, texts, properties, categories, brands, product images and similar.
• No customer data and no order data. plinth checks every request before it is sent. Requests with fields that are not product data are refused, and so are texts that contain e-mail addresses, bank account numbers or phone numbers.
• Results are always suggestions in the review list, never direct changes. A person accepts or rejects them.
• If a feature is switched off or there is no key, nothing is sent.

How do I set up the spell check?

The section Spell check in the AI settings connects a LanguageTool server of your own. It is off by default.

1. Switch on Turn on spell check and enter the Server address.
2. Under More settings: language, ignored words, rules you can add Ignored words, one word per line: brands, material names, custom terms.
3. Click Save.

Only product texts and property names are sent to that server. If the server can't be reached, everything continues without the check.

Common errors and fixes

See also

• Products
• Setup
• Onboarding wizard
• Access
• Quality