Parameters
Tune generation quality, diversity, and budget with OpenAI-compatible parameters.
POST/v1/chat/completions
| Name | Type | Required | Description |
|---|---|---|---|
temperature | number (0-2) | Randomness control. Typical range: 0 to 1. | |
top_p | number (0-1) | Nucleus sampling probability mass. | |
max_tokens | integer (>=1) | Maximum output tokens for completion. | |
frequency_penalty | number (-2 to 2) | Penalizes repeated tokens based on frequency. | |
presence_penalty | number (-2 to 2) | Encourages topic diversity by penalizing seen tokens. | |
stop | string | string[] | Stops generation when one of the sequences is produced. | |
seed | integer | Best-effort deterministic sampling for reproducible outputs. |
Recommended Profiles
Deterministic
{
"temperature": 0,
"top_p": 1,
"max_tokens": 512
}request-body.json
{
"model": "anthropic/claude-sonnet-4.5",
"messages": [{"role": "user", "content": "Write a concise launch update."}],
"temperature": 0.2,
"top_p": 1,
"max_tokens": 240,
"presence_penalty": 0,
"frequency_penalty": 0
}Tune One Knob First
Prefer changing either
temperature or top_p first. Adjusting both aggressively can make generation behavior harder to predict.