Pricing
How many credits each model and each service uses, read live from the catalogue and the configuration.
Every request uses credits from your account; your credits are loaded by your administrator. The tier of the request multiplies the prices counted in tokens; audio has a fixed price per unit.
Chat models
Chat completions, completions, the Responses API and the answers of knowledge bases are counted in tokens, per model, with input and output counted separately: generating tokens costs far more than reading them. Credits per million tokens.
Input tokens cost the model's list price on every hardware. Output tokens cost the list price times the factor of the hardware that answers. Write model@hardware to choose the machine and pay its output price; with the model's name alone the service picks a machine that serves the model, at that machine's price. Each row below is one model on one kind of hardware.
A long prompt pays its input more, on the whole input of that request: over 32,768 tokens ×1.5, over 131,072 tokens ×2. The input prices below are for prompts up to 32,768 tokens.
| Model | Hardware | Input (credits / 1M tok) | Output (credits / 1M tok) |
|---|---|---|---|
gemma-4-26b@blackwell6000proGemma 4 26B |
NVIDIA RTX PRO 6000 Blackwell 🇨🇭 Switzerland |
0.345000 |
0.600000 |
ministral-3-14b@blackwell6000proMinistral 3 14B |
NVIDIA RTX PRO 6000 Blackwell 🇨🇭 Switzerland |
0.345000 |
0.600000 |
mistral-medium-3.5-128b@blackwell6000proMistral Medium 3.5 |
NVIDIA RTX PRO 6000 Blackwell 🇨🇭 Switzerland |
1.495000 |
6.500000 |
qwen2.5:1.5bQwen 2.5 1.5B |
Hardware not specified 🇨🇭 Switzerland |
0.057500 |
0.100000 |
An image sent to a model that reads images becomes input tokens, at the input price of the model.
Other services
| Service | Price | Notes |
|---|---|---|
EmbeddingsPOST /v1/embeddings |
0.120000 credits per million input tokens
|
Tokens are estimated at about four characters each, the same count you see in usage.
Multiplied by the tier.
|
RerankPOST /v1/rerank |
not charged | The candidates, plus the query once for each candidate, estimated as for embeddings. |
Knowledge bases: a questionPOST /v1/knowledge-bases/{slug}/ask, and the app API |
the chat model's prices | The prices of the chat model that writes the answer, above: the question and the passages are its input. The search itself is not charged. |
Knowledge bases: indexing a documentupload in the dashboard or the app API |
0.120000 credits per million tokens of text indexed
|
Multiplied by the tier. |
TranscriptionPOST /v1/audio/transcriptions |
0.000100 credits per second of audio
an hour of audio: 0.360000 credits |
The duration of the file, as the engine measures it. |
SpeechPOST /v1/audio/speech |
0.000005 credits per character of the text
1,000 characters: 0.005000 credits |
What is not charged
- A request refused before it runs: any
4xx, for a parameter, the key, a rate limit, credits, or a prompt refused by the content check. - A request that fails on our side: a
502or a503, a stream that breaks before its first piece of text, a question to a knowledge base whose documents could not be searched or that has no document ready. - A silent WAV file, answered with empty text without asking the model.
What is charged although it is incomplete: an answer cut at the token cap, with the tokens it used; a stream you close early, or that breaks after its text has started, with the tokens the engine reported or, when it did not, an estimate of about four characters per token; and with n above 1, every answer.
How a request is counted
For a request with P prompt tokens and C completion tokens, on a model with input price in and output price out per million tokens, at tier T:
cost = (P × in + C × out) / 1_000_000 × multiplier[T]
Cached prompt tokens (usage.prompt_tokens_details.cached_tokens) are charged at 90% of the input price.
Embeddings, rerank and indexing have input tokens only. Audio is counted in its own units, in the table above, and the tier does not change its price. What each request cost is in your usage records, in the dashboard.
Tiers
The tier (X-Siati-Tier: slow|medium|fast|ludicrous, or the key's default) decides which machines may serve the request, the rate limit of your key, and a multiplier on the prices counted in tokens. It does not change the order in a machine's queue. More in Tiers.
| Tier | Rate limit | Price multiplier | Description |
|---|---|---|---|
| slow | 60 req/min | 0.85× |
Coda condivisa, latenza non garantita. |
| medium | 120 req/min | 1.00× |
Equilibrio prezzo/velocita'. |
| fast | 240 req/min | 1.25× |
Coda prioritaria, alto rate limit. |
| ludicrous | 1000 req/min | 1.60× |
Coda a massima priorità, rate limit elevatissimo. |
| company | 3000 req/min | 1.00× |
— |