Reasoning

Reasoning models expose reasoning_effort (and reasoning detail in responses) where the underlying model supports it:

curl https://api.vichar.io/v1/chat/completions \
  -H "Authorization: Bearer $VICHAR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vichar-sonnet-5.5",
    "reasoning_effort": "high",
    "messages": [{"role": "user", "content": "Solve x^2 - 5x + 6 = 0 step by step."}]
  }'

• Reasoning content appears in the response as reasoning fields on the message or stream deltas.
• usage.completion_tokens_details.reasoning_tokens counts reasoning tokens; they are billed as output tokens.
• supported_parameters lists reasoning for models that accept it, and reasoning_effort where effort levels are tunable. Passing the parameter to a non-reasoning model is ignored or rejected with 400 depending on the model.
• To force non-reasoning candidates in auto routing, send no_reasoning: true; if no non-reasoning model can serve the request you get a 400.