Thinking Variant

Enable reasoning-first behavior with :thinking

The :thinking variant routes to endpoints configured for deeper reasoning behavior, making it a better default for complex planning, math, and multi-step analysis tasks.

json

{
  "model": "deepseek/deepseek-r1:thinking",
  "messages": [
    {
      "role": "user",
      "content": "Design a migration plan from monolith to microservices with risk controls."
    }
  ]
}

Reasoning budgets

For cost and latency control, pair :thinking with explicit token constraints and request-level SLAs for production workflows.

For additional guidance on trade-offs, seeReasoning Tokens.