Coinbase AI Agent Cuts Fraud by 30% in Onramp Experiment

The adapted model caught a larger share of fraudulent transaction value, a measure that can matter more than simply counting detected cases.

Coinbase has fine-tuned an open-weight AI model on its proprietary transaction data to outperform a leading closed-source model in fraud detection, with better detection scores and lower serving latency.

In a three-part technical series, the exchange argued that companies can get better results from specialized models trained for their own needs than from relying on increasingly capable general-purpose AI systems.

Coinbase Reports Better Fraud Detection With a Smaller Model

In an October 12 post, Coinbase CEO Brian Armstrong explained that the company has fine-tuned Qwen3.5-9B using internal data and found that it “outperformed frontier models, at much lower cost and speed,” on the specialized task of fraud prevention.

The intended point was not that smaller models are universally better, but that performance depends on the job they are trained to do.

“The frontier models are amazing generalists,” Armstrong wrote. “But increasingly you’ll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.”