Coinbase Fine-Tuned Open-Source Model Beats Frontier AI Systems at Fraud Detection

Coinbase revealed that a specialized 9-billion-parameter open-source model, fine-tuned on internal fraud data, outperforms industry-leading frontier AI systems at detecting payment fraud while operating at a fraction of the cost. In a three-part technical blog series published October 6–8, 2026, CEO Brian Armstrong and the engineering team detailed how Alibaba's Qwen3.5-9B model—adapted using reinforcement learning on proprietary Onramp fraud data—achieved a 9.6% higher F1 score than Anthropic's Opus 4.5, while processing requests 55% faster (0.683 seconds versus 1.515 seconds) and requiring less than $100 to train.