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

518 points · 123 comments on HN · read original →

Points and comments are a snapshot, not live.

A hybrid Jev+Opus model matches Opus accuracy at 4x lower cost for identity resolution.

Hightouch tested whether decision models (Jev, GPT-6 Luna) could judge identity resolution over-merges. On an easy benchmark, Luna (97.3%) and Jev (95.7%) approached Opus (99.0%) at 0.15-0.17s and $0.05-$0.10 per 1,000 identities vs Opus's $3.86. On harder cases, Luna fell to 88.8%, matching Jev (89.1%). Jev's confidence is better calibrated; most mistakes are low-confidence. Feeding Jev's uncertain cases to Opus (30-90% confidence range) achieved 95.3% accuracy on hard cases for $2.94 per 1,000, versus Opus alone at 95.1% for $5.39. Hybrid Luna only reached 91.6% due to overconfidence.

The hybrid approach dynamically scales spend with problem difficulty.

What commenters are saying

Commenters widely endorse the hybrid approach, sharing similar setups in security and code generation. Several recommend using small or local models (Qwen3.8-27B, Haiku) as orchestrators or subagents to reduce cost while maintaining quality.

Two contrasting patterns emerge: using a cheap model to filter for a frontier model (as in the article), or using a smart orchestrator to delegate to cheap workers. One commenter notes that Haiku 5 is often sufficient for simpler tasks, outperforming Qwen models in their experience. Others describe practical multi-level setups with Opus, GPT-6 Sol, and Muse Spark.