MODEL CAPABILITIES
Reasoning controls
Some models reason before they answer. There is no request parameter to set how much.
H1 reasons: it produces thinking tokens before the answer. Whether and how long a model reasons is the model's choice; the API has no reasoning-effort parameter.
What you control
max_tokens. Reasoning tokens count toward it. Give reasoning models room: 1024 or more.- The model. Use Z1 or Z1 Code for short answers, and H1 when you want the model to think first.
- The prompt. A short, direct instruction tends to need less thinking.
What happens when the budget runs out
If max_tokens is exhausted while a model is still reasoning, the response is HTTP 200 with an empty content and finish_reason: "length". You are charged for the tokens produced. Raise max_tokens and retry.
Reasoning text
- Streaming responses carry only the answer in
delta.content; reasoning is not streamed. - Non-streaming responses include
reasoningandreasoning_contentfields on the message. They are null unless a model produced reasoning text, and when present they hold that text. - These two fields are not part of the stable contract. Do not build features on them or show them to end users as Tessen output.
Note
Reasoning tokens are billed as output tokens at the model's output price, which is why a one-word answer from H1 can use well over a hundred output tokens (one measured request: 176 output tokens, $0.0054).