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.
More from coding & agent
- Astra's Context Compaction Takes 3x Longer Than Sol, Dev Notes — banteg · 2026-09-16
- Simfinity.js example: expose selected GraphQL operations as MCP tools via stdio — Affectionate_Hawk971 · 2026-09-16
- GPT-6 spent $367 and 6.1M tokens over 7h50m to build a rocket in The Universim on its own — Angaisb_ · 2026-09-16
- DataFast Adds MCP/API Support for Mention Tracking and Chart Notes — marclou · 2026-09-16
- Semantica Hits 13k Stars: Open-Source Graph-Native Infrastructure for Accountable AI — tom_doerr · 2026-09-16
- Dev says his session-less personal assistant with continuous memory has users hooked for months — matteoianni · 2026-09-16