Skip to main content

Reasoning Models

Reasoning models spend extra internal "thinking" tokens before answering, which improves hard tasks (math, code, multi-step planning) at the cost of more output tokens. Primex forwards reasoning controls to the provider unchanged.

Requesting more or less reasoning​

On models that advertise the reasoning capability (see the model's catalogue entry), send the provider-native control, for example reasoning_effort on /v1/chat/completions:

{
  "model": "YOUR_REASONING_MODEL",
  "reasoning_effort": "high",
  "messages": [{"role": "user", "content": "Prove that √2 is irrational."}]
}

For native /v1/responses, use "reasoning": {"effort": "high"} with input. Accepted effort values depend on the selected model and its supported parameters.

Agentic clients expose this as a setting, for example Codex CLI's model_reasoning_effort (see Client Setup → Codex).

Billing​

Reasoning tokens are counted by the provider as output tokens and billed at the model's normal output rate. There is no separate "reasoning" meter. Higher effort means more output tokens, so it costs more; size the effort to the task. See Billing & Credits.

Tips​

  • Reasoning models are usually overkill for short or formatting-only tasks; a smaller chat model is cheaper and faster there.
  • Combine with structured outputs when you need the final answer in a machine-parseable shape.