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

A recent study explored whether fast, cost-effective models like Jev and GPT-6 Luna could match the accuracy of leading LLMs such as Claude Opus in grading identity resolution tasks. Researchers discovered that combining Jev with Opus achieved frontier-level accuracy while reducing costs fourfold. Identity resolution, the process of linking data to corresponding individuals, is crucial for targeted outreach, personalization, and analytics.

However, the task is challenging due to factors like shared emails, address changes, and data errors. Traditional methods rely on human review, which is slow and imprecise. The study utilized pseudopeople, a tool that simulates a fictional town with realistic data errors, to build 1,000 identities, half of which were one person with noisy records and the other half were over-merges.

By adjusting the confidence thresholds of Jev and Luna, researchers found that the decision models achieved accuracy comparable to frontier models at a fraction of the cost. The findings suggest that leveraging decision models in conjunction with LLMs could significantly enhance the efficiency and affordability of identity resolution processes.