Customer data company Hightouch published a test on November 4 in which a cheap decision model, Jev, graded 1,000 test identities and handed only its unsure cases to Claude Opus 5.5 (Hightouch blog). The pair scored 98.8% accuracy against 99.0% for Opus alone, at $0.88 per 1,000 identities against $3.86.
By our arithmetic that is 4.4x cheaper. On a harder test the same setup scored 95.3% against Opus’s 95.1%, at $2.94 against $5.39 per 1,000, or 1.8x cheaper by our arithmetic. The post reached 566 points on Hacker News.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how a harness around models shapes what they deliver. It teaches the general idea only and does not discuss Hightouch, Jev or Luna.
The key insight: Hightouch reported on November 4 that a cheap decision model, Jev, which handed only its unsure cases to Claude Opus 5.5, scored 98.8% against 99.0% for Opus alone at $0.88 instead of $3.86 per 1,000 identities graded.
The task was grading identity resolution: deciding whether a set of records merged into one customer profile really describes one person. The post says human reviewers manage about 50 identities an hour, and that estimating a rare over-merge rate precisely takes more than 1,000 samples.
Hightouch built 1,000 test identities with pseudopeople, a simulation tool from the Institute for Health Metrics and Evaluation at the University of Washington. Half were one person with noisy records and half were over-merges, the post says. No real people appear in the data.
The cheap models were Jev and OpenAI’s GPT-6 Luna on the Decisions API, which the post describes as a new class of models that return a 0-1 confidence score that a condition is true. The frontier models were Claude Opus 5.5, Gemini 3.1 Pro and GPT-6 Sol.
Cost per 1,000 identities graded in Hightouch’s test: Claude Opus 5.5 alone (grey) against a cheap decision model that hands its unsure cases to Opus (teal), with each setup’s accuracy. Data: Hightouch blog, 4 Nov 2026, costs observed in its own setup. The 4.4x and 1.8x ratios are our arithmetic.
The Numbers
On the first test, Opus 5.5 scored 99.0% at $3.86 per 1,000 identities and 2.2 seconds each. Gemini 3.1 Pro and GPT-6 Sol each scored 98.3%, at $4.48 and $2.50. Luna scored 97.3% at $0.10 and Jev 95.7% at $0.05, each in under 0.2 seconds.
The post sums it up: “Decision models are up to 90× cheaper and more than 10× faster out of the box”. On the harder test, which used only tricky cases such as twins and a junior and a senior, Opus scored 95.1% while Jev fell to 89.1% and Luna to 88.8% alone.
How the Hybrid Works, per the Post
The hybrid setup sends a case to Opus only when the cheap model’s confidence falls in a middle band. With Jev, sending identities scored between 30% and 90% meant 22% of cases went to Opus. That gave 98.8% accuracy at $0.88 per 1,000, the post’s table shows.
The post found Jev’s confidence better calibrated than Luna’s: “Most of Luna’s mistakes come with high confidence”. The post links that overconfidence to the harder-test results: hybrid Luna, at $1.15 per 1,000, reached 91.6%, while hybrid Jev reached 95.3%. “Calibration matters!”, the authors wrote.
Hightouch said hybrid Jev sent more cases to Opus on harder data without any change to the harness, calling it “a system that will automatically scale spend when needed to maintain accuracy, which is what we want”.
The Business Lever, by Our Arithmetic
In this setup, the frontier-model bill rises with the share of cases the cheap model is unsure about. By our arithmetic, the hybrid’s discount to Opus alone shrank from 4.4x on the first test to 1.8x on the harder one, as the post says more cases were sent to Opus on the harder data.
As we read the post, the result rests less on the cheap model’s raw accuracy than on how well its confidence scores flag the cases it may get wrong.
As we read the tables, the hybrid turns the frontier model into an escalation path: Opus was called on 22% of identities in the first test, and the post says more were sent on harder data.
As we read it, the test is one task on synthetic data, and the post says its costs and latencies will vary outside its own setup.
Hightouch blog, 4 Nov 2026
“Hybrid Jev’s ability to handle arbitrary problem difficulty is very valuable.”
Three Implications
FOR BUYERS OF FRONTIER MODELS In Hightouch’s test, the share of cases escalated set the Opus bill: 22% escalated cost $0.88 per 1,000 against $3.86 for Opus alone.
FOR DECISION-MODEL MAKERS The post found Jev’s confidence better calibrated than Luna’s, and wrote that most of Luna’s mistakes come with high confidence.
FOR DATA TEAMS The post says human review runs at about 50 identities an hour, while the decision models answered in about 0.15 to 0.2 seconds each.
The guide says “The new job of finance is to manage the ratio between the two” and that “That ratio is the gross-margin equation of the agentic era.”
As we read it, Hightouch’s hybrid is one way to move that ratio: the expensive model is paid for only on cases the cheap one is unsure about. The guide does not discuss Hightouch, Jev or Luna.
What Is Not Established
This is one company’s test on one task, with synthetic data. The post says “Model costs and latencies are what we observed in our own setup and will vary”. We did not rerun it. The post does not test other tasks, open-weight decision models or other escalation targets, and it reports no confidence intervals. This piece makes no forecast.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
In Hightouch’s test, a cheap model that escalated only its unsure cases got within 0.2 points of Claude Opus 5.5 at $0.88 instead of $3.86 per 1,000 identities. On harder data the saving fell to 1.8x by our arithmetic, but accuracy held at 95.3% against 95.1%.
A note on sourcing. Figures and quotations are from Hightouch’s blog post of 4 November 2026, which reports costs observed in its own setup on synthetic data. We did not rerun the test. Ratios are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
