Hightouch Didn't Pitch a Product. It Published Exact Numbers Across Two Difficulty Tiers — and That's Why It Hit the HN Front Page.

The Play

Hightouch had a genuinely useful finding: a cheap model, routed correctly, can match a frontier model's accuracy at a fraction of the cost.

They could have written that sentence and stopped. Most companies do.

Instead they published the receipts:

• "99.0% accuracy, 2.2 s per identity, $3.86 per 1,000" for Opus alone
• "95.7% accuracy, 0.15 s, $0.05 per 1,000" for the cheap model alone
• "98.8% accuracy, $0.88 per 1,000" for the hybrid setup

Then they ran a second, harder test scenario and published those numbers too — where the hybrid setup actually beat Opus: "95.3% accuracy, $2.94 per 1,000" versus Opus's "95.1% accuracy, $5.39 per 1,000."

Why this matters

Vague performance claims are easy to write and easy to ignore. "Dramatically cheaper, barely less accurate" could mean almost anything, so a skeptical reader has no reason to engage past the headline.

Specific numbers do the opposite. They invite scrutiny, and scrutiny is what makes something spread in a technical community — the post reached Hacker News's front page (202 points, 43 comments) precisely because people could pull a number out of it and debate whether it held up.

What the sourcing actually did

Hightouch named the models being compared instead of describing them abstractly: Jev, GPT-6 Luna, and Claude Opus 5.5. That alone lets a reader who's used any of those models calibrate the claim against their own experience.

They also ran two test scenarios instead of one. A single "it worked" result reads as a demo. Two results, with the harder one showing the hybrid approach pulling ahead, reads as a finding that holds under pressure.

And they explained the "why" in one sentence anyone technical can parse: confidence calibration, not raw accuracy, is what decided the winner. That sentence is the part other people quote back.

The growth play to steal

If you want a technical result to travel past your existing audience:

1. Publish exact figures instead of rounded or directional claims

2. Name every model, tool, or competitor in the comparison

3. Test under more than one condition so the result survives a skeptical second look

4. Give the mechanism a one-sentence explanation, not just a scoreboard

5. Credit named authors so the post reads as research, not marketing copy

Why most teams skip this

Publishing exact numbers feels risky — what if a competitor's model looks better in six months, or a reader finds a flaw in the methodology.

But that risk is also what makes the post worth trusting. A number specific enough to be wrong is a number specific enough to be believed.

Bottom line

Hightouch's post didn't go wide because the cost savings were impressive. It went wide because the numbers were exact, the models were named, and the test ran twice. That is a distribution strategy, and it costs nothing but the discipline to publish the real figures instead of a rounded pitch.

Source: https://hightouch.com/blog/llm-judges-identity-resolution