Control reasoning/thinking tokens on OpenRouter reasoning models (unified reasoning parameter)

Documented steps

1. Add a reasoning object to the /api/v1/chat/completions body. Fields: effort ("low" | "medium" | "high", plus "minimal"/"xhigh"/"none" on some OpenAI/Grok models), max_tokens (token budget for thinking — Anthropic/Gemini style), exclude: true (think internally but omit reasoning from the response), enabled: true (turn on at default/medium effort).
2. Pick the right knob per family: OpenAI o-series/GPT reasoning models take effort; Anthropic takes reasoning.max_tokens (minimum 1024); Gemini thinking models map effort/max_tokens to their thinking budget.
3. Reasoning output comes back in a reasoning_details array on the message (entries typed text / summary / encrypted) alongside normal content.
4. The legacy include_reasoning boolean is deprecated: include_reasoning: false ≙ reasoning: {exclude: true}.
5. Audit actual thinking spend: GET /api/v1/generation?id=<gen id> → native_tokens_reasoning, total_cost.
6. Find models that support it: GET /api/v1/models?supported_parameters=reasoning.
7. Docs: https://openrouter.ai/docs/use-cases/reasoning-tokens