# Hightouch — a Jev that asks Opus for help matches Opus at 4x lower cost

> A cheap decision model handles most cases and asks a frontier model only when it is unsure.

- Org: Hightouch
- Date: 2026-11-04
- Category: article, benchmark
- Source: https://hightouch.com/blog/llm-judges-identity-resolution
- Canonical page: https://ai-tldr.dev/releases/hightouch-jev-asks-opus-identity-resolution/

## In short

Hightouch tested Jev and GPT-6 Luna as cheap judges of identity resolution. A Jev that sends unsure cases to Claude Opus 5.5 scored 98.8% against Opus's 99.0%, at $0.88 versus $3.86 per 1,000 identities.

## What it is

Hightouch's November 4, 2026 post tests whether fast decision models can grade identity resolution output as well as frontier LLMs. The task is spotting over-merged identities, where records from different people were joined into one. The team built a 1,000-identity benchmark with pseudopeople, a simulated-town data generator, plus a harder set with typos and near-identical people such as twins.

## How it works

Jev or GPT-6 Luna screens each identity and returns a 0–1 confidence score. Identities whose score falls in a middle band, 30–90% for Jev, go to Claude Opus 5.5 for a second opinion. On its own, Jev scored 95.7% for $0.05 per 1,000 identities on the easier test.

## Why it matters

The hybrid setup is a simple recipe for cutting judge costs without losing much accuracy. On the easier test it was about 4.4x cheaper than Opus alone. On the harder test the saving shrank to about 1.8x ($2.94 versus $5.39), but the hybrid Jev was slightly more accurate, 95.3% against 95.1%. The post does not link released code or data.

## Who it's for

data teams running LLM judges at scale

## Links

- [Hightouch Post](https://hightouch.com/blog/llm-judges-identity-resolution)
- [HN Discussion](https://news.ycombinator.com/item?id=50397429)
- [pseudopeople](https://www.pseudopeople.org/)