Two-pass routing tries a query with a cheaper model before deciding whether a stronger model is needed. If the first answer appears incomplete, incoherent, or low-confidence, the system escalates the query. This keeps routine requests on the cheaper path while giving difficult requests another chance. section 9.2.4 section 9.2.6
The decision happens after the first answer:
- Send the query to the cheaper model and inspect its response for adequacy. section 9.2.4
- Keep an adequate answer; otherwise, send the query to the stronger model. section 9.2.6
A borderline query
Try gpt-5.4-mini first. If its response has low confidence or coherence, escalate to gpt-5.4. A query whose first answer is adequate avoids the stronger-model call. section 9.2.6
This differs from routing solely on the incoming query: keyword rules, a classifier, or a smaller model using function calling can choose a destination before an answer exists. Two-pass routing instead uses the first answer as evidence. No specific confidence metric or escalation threshold is prescribed, so that decision needs its own evaluation. section 9.2.4 section 9.2.6
