Coinbase Qwen Model Cuts Fraud by 30%
Coinbase says its fine-tuned Qwen3.5-9B model improved fraud detection for its Onramp crypto-buying service, reducing fraudulent transactions by 30% and the dollar value of fraud by 22% in A/B testing. The Coinbase fraud detection agent reviews transactions after existing machine-learning checks and returns a risk rating. On a proprietary benchmark of 16,140 transactions, including 813 confirmed fraud cases, the model outperformed Anthropic’s Opus 4.5 across precision, recall, F1 score and dollar-weighted recall. It had a median response time of 0.683 seconds, 55% faster than Opus 4.5, and training cost less than $100 on a single GPU. Coinbase also said newer versions of several frontier models performed worse than their predecessors on the benchmark. The results suggest tailored AI models could help crypto platforms lower fraud losses and operating costs, although the proprietary test has not been independently verified and applies to one product and task.
Neutral
The development is positive for Coinbase’s operating efficiency, but its direct effect on crypto prices and market stability is likely limited. Better fraud screening could reduce chargebacks and losses, strengthening confidence in Coinbase’s Onramp service over time. However, the article gives no evidence of an immediate change in trading volumes, user demand, or the performance of any specific cryptocurrency. Like other exchange technology and security updates, the news may attract short-term attention but is unlikely to move broad crypto markets on its own. The benchmark is proprietary, covers one product-specific use case and may not reflect how the model performs as fraud tactics evolve. Traders are therefore more likely to view this as a modest long-term operational positive than as a market-wide bullish or bearish catalyst.