Coinbase's compact AI model outperforms larger rivals in fraud detection
Coinbase fine-tuned Alibaba's Qwen3.5-9B model for fraud detection on its Onramp service. The tuned model reportedly beat frontier systems including Anthropic's Opus 4.5 on a proprietary benchmark, with lower latency and a training cost under $100. A/B testing showed a 30% reduction in fraudulent transactions and a 22% drop in fraud dollar value.
Coinbase described an AI fraud-risk agent for its crypto-buying Onramp service. It operates alongside existing machine-learning filters, reviewing transactions those systems already approved and returning only a risk rating. In A/B tests, it reduced fraudulent transactions by 30% and fraud dollar losses by 22%; estimated savings were triple the model's operating expense.
The comparison used a private benchmark of 16,140 transactions, including 813 confirmed fraud cases. After reinforcement-learning fine-tuning, Qwen3.5-9B surpassed Anthropic's Opus 4.5 on precision, recall, F1, and dollar-weighted recall by 7.5 to 35.4 percentage points. Its median response time was 0.683 seconds versus 1.515 seconds, and training cost under $100 on one NVIDIA RTX PRO 6000 96GB GPU.
If replicated, cheaper narrow fraud models could let more crypto services deploy stronger screening, potentially reducing card theft and chargebacks for users. Fraudsters may adapt, so benefits could be temporary. Smaller exchanges might benefit most because low training costs lower barriers. But proprietary benchmarks and single-product scope mean outside observers may not verify claims, so broader effects remain uncertain.