Token Types and Billing
A token is a text building block, roughly a word fragment. As a rule of thumb, about 1,000 tokens ≈ 750 words.
Four Token Types
Section titled “Four Token Types”| Token Type | Description | Billing |
|---|---|---|
| Input Token | Text the customer sends to the API (prompts, system instructions, RAG context) | Fully billable |
| Cached Input Token | Input tokens with an existing KV cache (prefix caching) | Billed at a reduced rate |
| Reasoning Token | Token for the thinking process (chain-of-thought), delivered to the customer | Fully billable (same rate as output) |
| Output Token | Response generated by the model | Fully billable (higher rate) |
KV Cache Prefix Caching
Section titled “KV Cache Prefix Caching”The KV cache resides volatilely in VRAM/RAM. It is not a persistent cache, the states are released as soon as processing completes. See Prefix Caching for details.
AI-Punkte
Section titled “AI-Punkte”Billing is usage-based (pay-as-you-grow) in AI-Punkte. A precommit is possible and economically attractive.
The specific AI-Punkte per model and an interactive consumption calculator are available in the AI-Punkte calculator.
Embedding and reranker models (Harrier, BGE) are billed exclusively via input tokens; cached input and output do not apply. Details on the models are available in the model overview.