Coinbase said a fine-tuned version of Alibaba's Qwen3.5-9B model outperformed frontier AI models at detecting fraud on its Onramp service. The company said training cost under $100. The benchmark is proprietary, so outsiders cannot independently verify the results.
Coinbase says a small AI model it fine-tuned now beats frontier models at catching fraud on Onramp, its service for buying crypto. In its "Owning Intelligence" blog series, published October 7-8, 2026, the exchange said the tuned version of Alibaba's Qwen3.5-9B was also faster and far cheaper to train.
A fraud agent that returns only a risk level
The project centers on an LLM-based fraud risk agent built for Onramp. The agent sits on top of Coinbase's existing machine-learning models and reviews transactions that have already cleared those systems.
Its output is narrow: it returns a risk level and nothing more, so upstream systems keep running as before. In A/B testing, Coinbase said the agent cut fraudulent transactions by 30% and lowered the dollar value of fraud by 22%.
Qwen against Opus 4.5
To compare models, Coinbase built a proprietary Onramp fraud benchmark of 16,140 transactions, 813 of them confirmed fraud. It fine-tuned Qwen3.5-9B with reinforcement learning, then tested it against frontier systems, including Anthropic's Opus 4.5.
The 9-billion-parameter model came out ahead on precision, recall, F1 score, and dollar-weighted recall, with margins of 7.5 to 35.4 percentage points. It also answered faster: median latency was 0.683 seconds, against 1.515 seconds for Opus 4.5, a 55% reduction.
Coinbase said training the model cost under $100 on a single NVIDIA RTX PRO 6000 96GB GPU.
Newer frontier models scored lower
Coinbase also reported that Opus 5, Sonnet 5, and GPT-5.6 underperformed their own predecessors on the same benchmark. The company appears to be leaning toward post-trained models tailored to narrow problems rather than general-purpose models.
The caveats matter, however. The findings apply to one task on one product, and fraud patterns shift as attackers adapt.
Source: Crypto Briefing
Trading involves risk.
Risk Disclaimer
