All news is rigorously fact-checked and reviewed by leading blockchain experts and seasoned industry insiders.
- Coinbase’s customized Qwen3.5-9B outperformed Claude Opus 4.5 across four fraud-detection metrics.
- The largest improvement involved detecting the monetary value of fraudulent transactions.
- Coinbase trained the model using reinforcement learning and proprietary transaction data.
- The results support using smaller, specialized AI models for narrowly defined financial tasks.
Coinbase has developed a specialized artificial intelligence model that outperformed Anthropic’s Claude Opus 4.5 in internal fraud-detection tests, strengthening the case for smaller AI systems trained on proprietary financial data.
The cryptocurrency exchange customized Alibaba’s open-weight Qwen3.5-9B model to assess transaction risk within Coinbase Onramp, its service for purchasing cryptocurrency through third-party applications.
The results, published in an October 8 engineering report, showed improvements in fraud identification and request-processing speed.
Coinbase CEO Brian Armstrong recently highlighted the research in X, suggesting that companies could increasingly deploy specialized models instead of relying exclusively on large general-purpose AI systems.
The findings illustrate a practical advantage of domain-specific training, although they do not establish that smaller models outperform frontier AI across broader applications.
Coinbase’s Qwen Model Beats Opus 4.5 on Fraud Detection
Coinbase evaluated the customized model against Claude Opus 4.5 and the original, unadapted Qwen3.5-9B.
The benchmark covered 16,140 historical transactions involving 7,293 users, including 813 confirmed fraudulent transactions collected over nine weeks.
Researchers used the same transaction cohort and decision framework to compare how accurately each model classified suspicious activity.
The post-trained Qwen model outperformed Opus 4.5 across all four detection metrics.
Its strongest result was in dollar-weighted recall, which measures the proportion of fraudulent transaction value successfully identified.
This distinction is important for payment providers. Detecting a larger share of high-value fraud can reduce financial exposure even when improvements in the overall number of fraudulent transactions detected are smaller.
The median latency result represents a 55% reduction in request time.
The P99 measurement is also operationally relevant because it captures performance near the slowest end of successful requests.
For a payment-risk system, unpredictable delays can be particularly problematic when customers are waiting for transactions to be approved.
However, these measurements cover the AI request itself, not the entire cryptocurrency checkout process.
How Coinbase Trained Its Nine-Billion-Parameter Model
Coinbase did not build a new foundation model.
Instead, it adapted Qwen3.5-9B using reinforcement learning with verifiable rewards, a technique that evaluates model outputs against predetermined outcomes.
In this case, the model learned to classify transaction risk using historical fraud labels rather than relying on another AI system to judge its responses.
The implementation combined several technologies:
- SkyRL: Managed the reinforcement-learning process and reward evaluation.
- Group Relative Policy Optimization (GRPO): Updated the model using comparisons between generated responses.
- Low-Rank Adaptation (LoRA): Reduced training requirements by updating a limited number of parameters.
- Ray Serve LLM and vLLM: Supported model deployment and online inference.
According to Coinbase, the training workload ran on a single NVIDIA RTX PRO 6000 GPU with 96GB of memory.
The company reported a compute cost below $100.
That figure excludes the broader expense of collecting transaction data, developing the fraud-detection system, engineering the training process and maintaining production infrastructure.
For deployment, Coinbase used NVIDIA L40S GPUs with autoscaling between two and eight replicas.
The company has not disclosed a comprehensive operating-cost comparison between its self-hosted Qwen implementation and equivalent commercial AI services.
Why Newer AI Models Did Not Always Perform Better
Coinbase’s decision to customize Qwen followed an earlier investigation into the reliability of commercial AI model upgrades.
In its October 7 engineering analysis, the company compared different versions of Claude Opus, Claude Sonnet and OpenAI’s GPT models using the same fraud benchmark.
Newer releases did not consistently produce better results.
For example, Coinbase found that a newer GPT model improved precision but reduced recall and dollar-weighted recall.
That means the model became more selective about identifying fraud while missing a greater share of fraudulent activity.
For financial risk management, this can create an unfavorable trade-off.
A model that generates fewer false alarms may appear more accurate in some circumstances, yet allow more fraudulent payments to proceed.
The findings encouraged Coinbase to evaluate AI models against specific financial outcomes rather than general-purpose performance rankings.
Customizing an open-weight model also gave the company greater control over its training objectives, model versions and deployment environment.
The Earlier Fraud Agent Had Already Reduced Losses
The Qwen project builds on Coinbase’s earlier work integrating an AI-powered risk agent into its payment infrastructure.
As detailed in the company’s October 6 engineering report, the agent reviews selected transactions after conventional machine-learning models and predefined rules perform their initial checks.
Its role is to identify suspicious patterns across recent transaction activity that may be difficult to detect from an individual payment.
In an online A/B experiment, the agent-assisted system recorded 30% fewer fraudulent transactions and 22% less fraudulent transaction value than the existing controls alone.
These results demonstrate the benefit of adding contextual AI review to Coinbase’s risk infrastructure.
They do not measure the subsequent Qwen upgrade, which was evaluated separately.
The distinction matters because the first experiment tested the business impact of introducing an AI agent, while the later research compared which model should power that agent.
What Specialized AI Means for Financial Infrastructure
Coinbase’s research offers a useful example of how financial institutions can approach AI deployment.
Large general-purpose models are capable of handling diverse tasks, but their breadth does not guarantee optimal performance in specialized applications.
Fraud detection has characteristics that make targeted training particularly valuable: historical outcomes can be measured, decision criteria can be defined and mistakes have identifiable financial consequences.
A smaller model trained on those outcomes may deliver stronger results while requiring less computing capacity.
Self-hosting also gives companies more control over sensitive transaction data, software updates and operational dependencies.
Armstrong said Coinbase is developing internal tools that allow employees to initiate model customization workflows more easily.
He also suggested that some backend services could eventually be replaced by specialized AI-based systems.
Such changes would require ongoing monitoring. Fraud patterns evolve, and models trained on historical transactions can become less effective as criminal behavior changes.
Coinbase’s own engineering team emphasizes continued evaluation, model versioning and rollback procedures as essential parts of maintaining a production system.
The company’s results demonstrate that model specialization can outperform scale on a well-defined financial task.
The broader commercial question is whether that advantage can be maintained as transaction volumes, fraud patterns and operating requirements change.
Be the first to comment