Reasoning: effort to thinking budget

What it does

This transform maps the request reasoning effort level (low, medium, high, and optionally xhigh or max) to a numeric thinking-token budget for upstreams that take budgets instead of effort labels.

Reference

Property	Value
Type ID	`reasoning_effort_to_budget`
Phase	`request`
Scopes	Provider

When to use

Clients send OpenAI-style reasoning.effort but the upstream takes an Anthropic-style thinking budget.

Configuration

Option	Type	Required	Default	Description
`low`	integer	Yes	—	Thinking-token budget for effort `low`.
`med`	integer	Yes	—	Budget for effort `medium`.
`high`	integer	Yes	—	Budget for effort `high`.
`xhigh`	integer	No	—	Optional budget for effort `xhigh`.
`max`	integer	No	—	Optional budget for effort `max`.

Example rule

{
  "transform": "reasoning_effort_to_budget",
  "enabled": true,
  "phase": "request",
  "models": [
    "claude-*"
  ],
  "config": {
    "low": 1024,
    "med": 8192,
    "high": 24576
  }
}

This transform is Provider-scope only, because it changes upstream spend.