Thinking Budgets via Model Name Parentheses

Append (value) to the model name to control the thinking budget or reasoning effort. The proxy strips the parentheses before routing and applies the derived settings to the request.

Accepted Values

• (number): explicit thinking budget (provider-native tokens); clamped to the model’s supported range.
• (level): preset reasoning effort (case-insensitive):

Level	Approx. Budget	Purpose
minimal	512	Lowest cost reasoning
low	1024	Fast reasoning
medium	8192	Default reasoning depth
high	24576	Deep reasoning
xhigh	32768	Extra-deep reasoning
auto	Dynamic (-1 when allowed, else mid/min)	Let the provider choose
none	0 (clamped to min when zero is disallowed)	Disable thinking

• Empty () is ignored. For provider://model forms, place the parentheses after the model (e.g., openrouter://gemini-3-pro-preview(high)).

How It’s Applied

• Only models that advertise thinking keep these settings; unsupported models simply drop the suffix without injecting thinking fields.
• Gemini (standard & CLI): writes generationConfig.thinkingConfig.thinkingBudget (or request.generationConfig.thinkingConfig... for CLI) after clamping. include_thoughts is left unchanged. Models with default thinking (e.g., gemini-3-pro-preview) still auto-enable thinking when missing; the parentheses budget overrides the default.
• Claude API: when a budget/level is provided, sets thinking.type=enabled with normalized thinking.budget_tokens and bumps max_tokens if needed.
• OpenAI/Codex/Qwen/iFlow/OpenRouter: reasoning levels/auto/none overwrite reasoning_effort (chat) or reasoning.effort (Responses). Numeric budgets do not change reasoning_effort for these protocols.
• Level-based models enforce their supported effort levels; unsupported values return HTTP 400.

Examples

• Dynamic budget with Gemini:

curl -X POST http://localhost:8317/v1/chat/completions \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{
        "model": "gemini-3-pro-preview(auto)",
        "messages": [{ "role": "user", "content": "Summarize the key points" }]
      }'
# Normalizes to gemini-3-pro-preview and sets thinkingBudget=-1 (clamped if dynamic is not allowed); include_thoughts stays unchanged.

• High reasoning effort for Responses:

curl -X POST http://localhost:8317/v1/responses \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{
        "model": "gpt-5.1(high)",
        "input": "List three improvements"
      }'
# Routes as gpt-5.1 and overwrites reasoning.effort="high".

• Disable thinking (clamped when zero is not allowed):

model=claude-sonnet-4.5(none)
# Sets thinking.budget_tokens to 0 when allowed; otherwise clamps to the model minimum.