Hightouch Hybrid Model Achieves Near-Opus Accuracy at 4x Lower Cost

• Hybrid Jev system achieves 95.3% accuracy on hard identity resolution tests, matching Claude Opus 5.5's 95.1% while reducing cost from $5.39 to $2.94 per 1,000 identities.
• Decision models Jev and GPT-6 Luna are up to 90x cheaper and 10x faster than frontier models, with Jev demonstrating superior confidence calibration for routing uncertain cases.
• On easier tests, the Hybrid Jev system reaches 98.8% accuracy at $0.88 per 1,000, whereas standalone Opus 5.5 achieves 99.0% at $3.86 per 1,000.
• GPT-6 Luna shows higher initial accuracy but suffers from overconfidence on difficult cases, resulting in lower hybrid performance (91.6% accuracy) compared to Jev-based routing.

Performance Metrics

• Harder Test: Hybrid Jev (95.3% acc, $2.94/1k, 1.56s) vs. Opus 5.5 (95.1% acc, $5.39/1k, 2.6s).
• Easier Test: Hybrid Jev (98.8% acc, $0.88/1k, 0.63s) vs. Opus 5.5 (99.0% acc, $3.86/1k, 2.2s).
• Standalone Jev (60% cutoff) costs $0.05-$0.06 per 1,000 with ~0.15s latency.

Methodology and Limitations

• Evaluation used synthetic US town data generated by the 'pseudopeople' tool, comprising 1,000 identities with 50% noise and 50% over-merge cases.
• Hybrid routing sends cases within specific confidence ranges (e.g., 30-90% for Jev) to Opus for re-evaluation, dynamically adjusting spend based on data difficulty.
• Reported costs and latencies are specific to Hightouch's internal environment and may vary in other deployments.