Coinbase's fine-tuned Qwen3.5-9B beats Opus 4.5 on fraud detection

Coinbase post-trained open-weight Qwen3.5-9B on its own Onramp fraud data using RLVR with GRPO and LoRA on SkyRL, for under $100 on a single RTX PRO 6000 96GB. On its internal benchmark the model beat Claude Opus 4.5 by +7.5 precision, +12.0 recall, +9.6 F1 and +35.4 dollar-weighted recall. In a 50% traffic A/B test the agent cut fraudulent transactions by 30% and fraud value by 22%.

• Fine-tuning on one RTX PRO 6000 96GB cost under $100
• Gains over Opus 4.5: +12.0 recall and +35.4 dollar-weighted recall
• A/B test: 30% fewer fraud transactions and 22% less fraud value
• P50 latency 0.683 s versus 1.515 s for Opus 4.5