Vaibhav Sisinty (@VaibhavSisinty) on X

Coinbase fine-tuned Qwen 3.5-9B, a 9 billion parameter open-source model, on their internal fraud detection data and got better results than Claude Opus 4.5 on their private benchmark. The whole operation cost under $100 and ran on a single GPU. They cut serving latency in half—from 1.515 seconds down to 0.683 seconds. When they deployed a fraud detection agent combining this model with their existing ML systems and rules, they caught 30% more fraudulent transactions and prevented 22% more fraud value. The savings from faster inference alone were roughly 3x the cost of running the model. The real story here isn't that a small model beat a frontier model in general intelligence. It's that specialized models trained on your own data will outperform general-purpose AI on narrow, high-stakes problems. Coinbase noticed something counterintuitive: newer frontier releases actually scored worse on recall and F1 for fraud detection specifically. That matters because those frontier models were trained on broad internet data, not on the particular patterns of fraudulent transactions in their system. A model trained on your company's actual problems will naturally beat something built to be good at everything. Coinbase is now building internal infrastructure to make this pattern repeatable. They're giving employees tools to fine-tune models in-house, including a command in their internal system that lets someone request something like "/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review." Brian Armstrong, Coinbase's CEO, expects this to become standard practice—companies spinning up specialist models for specific backend tasks instead of relying entirely on expensive frontier models. Some internal services might get replaced entirely by this new architecture.