Using confidence-based routing between GPT-5.5 and GPT-5.6 for product classification - worth it?

A developer is exploring confidence-based routing between GPT-5.5 and GPT-5.6 for classifying supermarket products into a multi-layer taxonomy. Initial tests on batches of 50 records showed GPT-5.6 to be noticeably more accurate than GPT-5.5. The user is undertaking this as a personal project and aims to minimize costs.