Confidence-based routing between GPT-5.5 and GPT-5.6: is it worth it

Expensive_Break_6163 · reddit · 2026-09-16

A developer tested GPT-5.5 (low) vs GPT-5.6 (medium) on supermarket product classification across a multi-layer taxonomy (e.g., Fruits → Frozen/Fresh/Canned), on batches of 50 records. GPT-5.6 was noticeably more accurate. He's considering a cost-saving routing scheme: let 5.5 handle most classifications and fall back to 5.6 only when confidence is low — but questions whether model self-reported confidence is reliable enough for that, on a minimum-cost personal project.

Original post →

More from coding & agent

coding & agent channel →