Model Cascading

Hightouch benchmarked its decision models Jev and Luna against frontier model Opus on identity-resolution grading. A hybrid harness that routes Jev's low-confidence cases to Opus reaches frontier-level accuracy at over 4x lower cost. The study also finds Jev's confidence is better calibrated than Luna's, while decision models remain up to 90x cheaper and 10x faster but less accurate on their own.