Coinbase has fine-tuned an open-weight AI model using internal transaction data and found it outperforms larger closed-source models on fraud detection while operating at lower cost and faster speeds, according to an October 12 post by CEO Brian Armstrong.
The company adapted Qwen3.5-9B, a smaller model, for its Onramp service, which allows users to purchase cryptocurrency through partner applications. In an online experiment, the AI agent identified 30% fewer fraudulent transactions and 22% less fraud value compared to transactions processed without the additional AI review.
How the System Works
Coinbase integrated an AI risk agent into its existing machine-learning model and rules engine. The agent examines sequences of recent transactions to identify fraud patterns that may not be visible from individual payments alone. The system can block a transaction that passed initial checks but cannot reverse an upstream block that has already been made.
Performance Comparison
When tested against a fixed benchmark of 16,140 transactions from 7,293 users—including 813 confirmed fraudulent transactions—newer versions of closed-source models showed mixed results. GPT-5.6 improved precision but achieved lower recall, identifying a smaller share of fraudulent transactions overall. These results prompted Coinbase to transition to an open-weight model whose parameters could be directly adapted.
Armstrong emphasized that the advantage of specialized models lies in their specific training rather than universal superiority. "Frontier models are amazing generalists," he wrote, "but increasingly you'll see specialist models outperform on more narrow tasks."
Safety Guardrails
Coinbase's agent operates within defined limits. It returns a risk score and relies on application code to apply the decision policy. Security firm CertiK has cautioned that autonomous AI agents in financial crime prevention require clear boundaries, including limits on autonomous decisions and designated accountability for each agent.
