Coinbase has developed an AI-based fraud detection agent for its Onramp service using a fine-tuned version of Alibaba’s Qwen3.5-9B model. The agent operates alongside existing machine-learning systems to assess transaction risk levels.
The company evaluated the model against a proprietary benchmark of 16,140 transactions, including 813 confirmed fraud cases. The 9-billion-parameter model outperformed frontier systems like Anthropic’s Opus 4.5 across precision, recall, F1 score, and dollar-weighted recall metrics, with margins between 7.5 and 35.4 percentage points.
The Qwen-based agent achieved a median latency of 0.683 seconds, a 55% reduction compared to the 1.515 seconds recorded by Opus 4.5. Training the model cost under $100 on a single NVIDIA RTX PRO 6000 96GB GPU.
A/B testing showed that the agent reduced fraudulent transactions by 30% and lowered the dollar value of fraud by 22%. Coinbase estimates that the savings generated are three times the cost of running the model.
Coinbase noted that newer frontier models, including Opus 5, Sonnet 5, and GPT-5.6, performed worse than their predecessors on the company's internal benchmark.
