Coinbase: Newer AI Models Underperform Predecessors at Fraud Detection

• —Benchmark covered 16,140 transactions from 7,293 users, including 813 confirmed fraudulent cases
• —Sonnet 5 recall fell 22.2 points versus Sonnet 4.6; dollar-weighted recall dropped 22.9 points
• —GPT-5.6 precision rose 11.5 points but recall fell 20.7 points
• —A post-trained Qwen3.5-9B beat Opus 4.5 on all four fraud metrics and cut serving latency by 55%