Coinbase's specialized Qwen3.5-9B model outperforms Claude Opus 4.5 on fraud detection

Coinbase has developed a specialized fraud detection model, Qwen3.5:9b, which reportedly outperformed Anthropic's Claude Opus 4.5 on Coinbase's internal fraud benchmark. The system integrates the LLM as a downstream agent that reviews selected transactions, rather than replacing existing ML models and rules engines. Key to their approach is a "tier-only" output for risk assessment and a "block-only" authority for the agent, ensuring it cannot approve transactions but can flag suspicious ones. This system, when A/B tested, reduced fraudulent transactions by 30% and fraud value by 22%, with an estimated fraud savings of three times the inference cost. AI

IMPACT Demonstrates the effectiveness of specialized, fine-tuned models for niche tasks and highlights a replicable system architecture for fraud detection.

RANK_REASON The item details a specialized model's performance on a specific benchmark, including methodology and results, which falls under research. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Akshit Trehan
• Aman Choudhary
• Anyscale, Inc.
• Bo Lei
• Brian Armstrong
• Claude Opus 4.5
• Coinbase
• Eric Tang
• Jess Kong
• Qwen3.5:9b
• Seiji Eicher
• Sumanth hegde
• X
• Xuwei Tan
• Yao Ma

• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →