A Jev that can ask Opus for help gives Opus accuracy at 4x lower cost

Researchers tested whether cheaper, faster decision models like Jev could match frontier LLM accuracy for identity resolution tasks and found that combining Jev with Claude Opus achieves comparable accuracy at over 4x lower cost. While Jev and Luna decision models trail Opus by about 6 percentage points on harder benchmarks when used alone, Jev's better-calibrated confidence scores make it effective in a hybrid model that delegates uncertain cases to Opus. Decision models offer significant efficiency gains—costing $0.5-$0.10 per 1,000 identities with 0.15-second response times—making them practical for cost-sensitive applications despite their accuracy gap.