Pricing and data
Pricing and prompt caching
What a request costs and how to pay less.
You pay for tokens: what you send (input) and what the model writes (output). Prices are in Danish kroner per 1 million tokens, excluding VAT, and are under Oversigt (Overview) on the platform. Each model has three prices:
| Price | Applies to |
|---|---|
| Input | Prompt tokens that weren't in the cache. |
| Cached input | Prompt tokens read from the cache. Much cheaper than input. |
| Output | The answer's tokens. |
The price of a request is
(prompt_tokens − cached_tokens) × input price
+ cached_tokens × cached input price
+ completion_tokens × output priceThe numbers are in the answer's usage. The price is taken from your balance when the answer is done. You can see your usage per day on the platform under Forbrug (Usage).
Balance and payment
You pay in advance. Under Betalinger (Payments) on the platform you buy credit by card, from 50 kr. at a time. Amounts exclude VAT. Danish customers pay 25% VAT on top, and businesses in other EU countries can enter their VAT number so no VAT is added. You get an invoice for each payment, which you also find under Betalinger.
Credit you buy can be used for 12 months from the purchase, and the oldest credit is used first. Expired credit is not refunded. See the terms, section 7.5.
The balance has a cap, shown under Betalinger. Write to info@gavn-software.dk if you need more, and we'll arrange it with an invoice.
When the balance is 0 kr. or below, the API refuses new requests with 429 and insufficient_quota. See Errors.
Prompt caching
Caching happens by itself. When a conversation continues, its start is the same as last time: the system prompt and the earlier messages. That part is already in the cache and costs the cached input price. Requests in the same conversation go to the same server, so the cache is hit.
To get the most out of it:
- Put what doesn't change first, such as a fixed system prompt.
- Add new messages at the end, and leave the earlier ones as they were.
The cache is split per account. Another account can never hit your cache.
Interrupted answers
| What happens | What you pay |
|---|---|
| You stop a stream along the way | The prompt and what the model managed to write. |
| You cancel a request without streaming | The whole answer. The model finishes it anyway. |
| The model stops answering | Nothing. |
The request is refused (for example 429 or 503) | Nothing. |
When you stop a stream, we can't yet see how much of the prompt was in the cache. Until we can, you pay the input price for the whole prompt in that case.