I opened Coinbase's three-part write-up on its in-house fraud model expecting a GPU story. The cheap training run is real, by their account. But the line I keep rereading is about something else: where the model's grades came from.
What Coinbase says it did
In Part 3, published October 8, 2026, Coinbase says it post-trained Qwen3.5-9B, an open-weight model, with reinforcement learning on verifiable rewards. In plain words, the model answers, and its answer gets scored against a known right answer. They used LoRA adapters through SkyRL on Anyscale, on one RTX PRO 6000 96GB card, for under $100 of training compute, by Coinbase's own account [1].
The result they report: the tuned 9B beat Opus 4.5 by 9.6 F1 points on Coinbase's private benchmark. That comparison is against Opus 4.5 only, on their test set, not a public one. I can't rerun it, and you can't either. Treat it as one company's report, not a law of nature.
Two more caveats, so nobody quotes this wrong. The "30% fewer fraud transactions" result in Part 1 came from the agent running on Opus, not from the 9B [2]. And Coinbase doesn't describe a live A/B test of the 9B in production.
The part that isn't $100
Here's the sentence that changed what I think this story is about. Coinbase scored the rewards against "known historical fraud outcomes" [1]. No AI judge grading the AI. Every training answer was checked against what actually happened: this transaction turned out to be fraud, that one didn't.
That's the answer key. And you can't buy it with a GPU rental.
Fraud is a good fit for this because it usually resolves itself over time. Chargebacks land. Accounts get confirmed. The truth tends to show up in a company's own records, case by case. What the write-up doesn't say is who confirmed those labels, or how long it took. That cost is undisclosed, so I won't guess it.
What I can give you is a floor for your own math. Noah Solace on our desk pulled the US Bureau of Labor Statistics median wage for data entry keyers, $19.88 an hour [3]. That isn't what Coinbase paid. It's a floor for what an hour of careful labelling costs if you hire it out. Bring your own seconds per label, because nobody has published a solid figure for that. Rex Circuit's Builder Lab piece shows how he modelled it.
What this means if you're small
The headline reads like "anyone can train a specialist for $100." The honest version is narrower: anyone who already holds a pile of confirmed answers for one narrow decision might.
That's actually good news for a lot of founders. You don't need Coinbase's scale. You need a decision that's:
- Narrow. The output is one choice from a short fixed list: fraud or not, refund or not, which queue a ticket goes to.
- Self-grading. The right answer shows up later without anyone arguing about it: a chargeback, a refund, a reopened ticket, a cancelled plan.
- Already logged. Those outcomes sit in your database next to the original case, so you can hold some back to test on.
- Owned. You know who confirmed each answer, and whether they'd agree with each other.
If one of those is missing, the cheap training run doesn't save you. You'll spend the money on building the answer key instead. That's still a fine project. It's just a different project from the one the headline sells.
And notice the people in it. Someone confirmed each of those outcomes. Coinbase doesn't say who, but their judgment is the asset the model learned from. If you're about to do this, the person who closes your tickets is your most important collaborator, not your GPU vendor.
Try this in the next 48 hours
Take ten minutes with whoever closes your support or ops cases. Pick one decision they make every day. Ask four questions: Is the answer one of a few fixed options? Does the true outcome show up later on its own? Is it logged next to the original case? Who confirmed it?
Four yeses means you may already own an answer key. Then time a few of your own labels and do the labelling math before you spend a dollar on GPUs. If you want to measure a free baseline first, Rex's Specialist Bake-off Kit is free for members. Any no is your first task, and it's cheaper to learn that now.
References
- Coinbase, "Owning Intelligence, Part 3," October 8, 2026. coinbase.com/blog/part-three-owning-intelligence-series
- Coinbase, "Part 1: How Coinbase built, evaluated and post-trained a fraud agent," October 6, 2026. coinbase.com/blog/part-one-how-coinbase-built-evaluated-and-post-trained-a-fraud-agent
- US Bureau of Labor Statistics, OEWS profile for Data Entry Keyers (43-9021), current release (May 2025). data.bls.gov/oesprofile
