Coinbase has completed development of a specialized artificial intelligence system designed to catch fraudulent transactions on its Onramp service for purchasing cryptocurrency. The exchange reported that a fine-tuned version of Alibaba's Qwen3.5-9B model outperformed larger frontier models in fraud detection while offering significant speed and cost advantages.
The fraud risk agent operates on top of Coinbase's existing machine-learning infrastructure and evaluates transactions that have already passed through those systems. The agent produces a narrow output—a risk level assessment only—allowing upstream systems to continue operating unchanged.
Performance Results
Coinbase conducted A/B testing on the system with measurable outcomes. The agent reduced fraudulent transactions by 30 percent and lowered the dollar value of fraud by 22 percent. The company estimated the savings at three times the cost of running the model.
To evaluate the system, Coinbase created a proprietary benchmark containing 16,140 transactions, of which 813 were confirmed fraud cases. The fine-tuned Qwen3.5-9B model, trained using reinforcement learning, was compared against frontier systems including Anthropic's Opus 4.5.
The 9-billion-parameter model exceeded frontier alternatives across four performance metrics: precision, recall, F1 score, and dollar-weighted recall. Performance margins ranged from 7.5 to 35.4 percentage points. Median latency for Qwen was 0.683 seconds, compared with 1.515 seconds for Opus 4.5—a 55 percent reduction.
Training the Qwen model cost under $100 on a single NVIDIA RTX PRO 6000 96GB GPU.
Broader Implications
Coinbase's results suggest that specialized, post-trained models tailored to narrow problems may outperform general-purpose frontier models for specific tasks. The company's approach prioritizes domain-specific optimization over reliance on the latest general-purpose model releases.
The findings carry limitations. The benchmark is proprietary, preventing independent verification of results. The performance improvements apply specifically to one product feature, and fraud patterns shift as attackers adapt strategies. Additionally, newer frontier model releases—including Opus 5, Sonnet 5, and GPT-5.6—underperformed their predecessors on Coinbase's benchmark, a pattern warranting further examination as models evolve.
