Cheap model first, escalate to Opus when unsure: same accuracy at half the cost
michaelTM_ai · reddit · 2026-11-08
A Hightouch cascade test (#2 on HN) routes a cheap confidence-scoring model first and escalates unsure cases to Opus for identity-resolution checks. Easy set: Opus alone 99.0% at $3.86/1k vs Jev+Opus 98.8% at $0.88. Hard set (twins, same names): Opus alone 95.1% at $5.39 vs Jev+Opus 95.3% at $2.94. Key insight: Luna is more accurate than Jev solo but loses in the cascade — its errors come with high confidence so the router never catches them. For cascades, a cheap model that knows when it's unsure beats one that's right more often.
More from coding & agent
- Viral prompt makes Claude Code build its own skills from your repeated workflows — Aiden_Tech_Ai · 2026-11-08
- Alibaba's Qoder Launches Stealth Model Qoduck, Free Daily for Individual Users — Aiden_Tech_Ai · 2026-11-08
- DECIDE benchmark tests voice-agent decisions on 4,867 questions from 1,081 real calls, scoring 20 models — Aiden_Tech_Ai · 2026-11-08
- Karpathy builds TristanTube in 3 minutes with Claude Code: 47K curated nerdy videos — mhdfaran · 2026-11-08
- Coinbase CEO runs always-on lead agents with one queue; skeptics note backlog risk — petrusenko_max · 2026-11-08
- Building OneContext on Hindsight: Shared Team Memory for AI Coding Assistants — kumar_2005 · 2026-11-08