OpenRouter guide: model tiered escalation routing using confidence thresholds
OpenRouter releases a tutorial on how to let cheap models return a confidence field from 0 to 1 via structured output, and route low-confidence requests to a stronger model. The article emphasizes that the confidence score is only self-reported, not a calibrated probability, and thresholds should be set based on the error rate by score band of your own traffic, and after launch monitor the score distribution, escalation rate and error rate of non-escalated answers and continuously adjust.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original