The Pulse

Hightouch’s Jev–Claude Opus Hybrid Nearly Matches Opus Accuracy at Lower Cost in Identity Resolution

In Hightouch’s tests of Jev and GPT-6 Luna as judges for identity resolution, with selected cases then routed to Claude Opus, the Jev-Opus hybrid scored 98.8% at $0.88 per 1,000 identities on the easier test, against 99.0% at $3.86 for Opus

AI.info Team · Written with AI · fact-checked · 2026-11-07

Hightouch reported on Nov. 4, 2026, that a system pairing Jev with Claude Opus approached Opus-alone accuracy in tests of identity-resolution judgments. On the easier test, the hybrid scored 98.8% accuracy at $0.88 per 1,000 identities; Claude Opus 5.5 alone scored 99.0% at $3.86. On the harder test, hybrid Jev scored 95.3% at $2.94 per 1,000, compared with Opus’s 95.1% at $5.39.

Hightouch’s 1,000-identity test

Identity resolution, as Hightouch defines it, is the process of mapping data to corresponding real people. Hightouch focused on over-merging: an identity that incorrectly combines records belonging to multiple people. Data issues include typos, shared emails, moved addresses and blank fields.

There is no answer key for checking identity-resolution results on real data, Hightouch writes, so teams sample profiles and inspect them manually. Human sampling is slow (~50 identities/hr), the company says, and > 1000 samples may be needed to get the error down to under +-1% on the over-merging rate. Hightouch used pseudopeople, a tool from the Institute for Health Metrics and Evaluation at the University of Washington, to simulate a fictional US town’s records and add realistic data errors.

The benchmark contained 1,000 identities: half represented one person with noisy records, and half were over-merges. No real people appear in any example, the post says. Models judged whether records in a profile described one person or several.

Jev’s confidence is better calibrated than Luna’s

At a 50% confidence cutoff, GPT-6 Luna scored 97.3% on the first test and Jev scored 95.6%; Claude Opus 5.5 scored 99.0%. Hightouch then tuned the cutoffs to 70% for Luna and 60% for Jev, producing scores of 97.3% and 95.7%, respectively. Luna performed better than Jev out of the box, but the gap disappeared on harder cases, the company says.

Most of Jev’s mistakes came with low confidence that it was right, while most of Luna’s came with high confidence. On the harder test, Jev alone scored 89.1% and Luna alone 88.8%; Claude Opus 5.5 scored 95.1%. Jev’s better-calibrated confidence, Hightouch says, helped the hybrid decide which cases to send to Opus for another judgment.

Opus reviews a selected share

In the easier test, Hightouch sent cases with Jev confidence between 30% and 90% to Opus; 22% of identities were sent. That hybrid scored 98.8% accuracy at $0.88 per 1,000 identities, compared with 95.6% and $0.05 for Jev alone. Sending Luna cases in the 10%–90% confidence range to Opus sent 6% of identities for review and scored 98.5% at $0.33 per 1,000.

On the harder test, the Jev hybrid scored 95.3% at $2.94 per 1,000 identities, while the Luna hybrid scored 91.6% at $1.15. Opus alone scored 95.1% at $5.39. The Jev system sent more cases to Opus as the data became harder, without any harness or prompting change, Hightouch says.

Observed costs vary with the test

Across the two tests, Jev alone cost $0.05 per 1,000 identities on the easier test and $0.06 on the harder one; Luna alone cost $0.10 and $0.12, respectively. The hybrid systems cost more when they sent more identities to Opus: hybrid Jev cost $0.88 on the easier test and $2.94 on the harder one. The model costs and latencies were observed in Hightouch’s own setup and will vary, the company says.