Coinbase CEO: AI Model Trained on Internal Data Shows High Fraud Detection Performance

[Block Media Reporter Ham Ji-hyun] Coinbase has recorded higher fraud detection performance using an open-weight AI model trained on its own payment data compared to general frontier models.

CEO Brian Armstrong stated on the 12th (local time) via X that "by fine-tuning the open-weight model Qwen 3.5-9B using Coinbase's proprietary internal data, we achieved much higher performance in the specialized task of fraud prevention at a significantly lower cost and faster speed than frontier models."

He added, "While frontier models are excellent general-purpose models, we will increasingly see specialized models outperforming in narrower tasks, either working alone or in collaboration with multiple models."

The speed difference was also evident in the online inference results shared by Armstrong. The median (P50) request processing time for Qwen 3.5-9B was 0.683 seconds, about 55% shorter than the 1.515 seconds for Opus 4.5.

In the 99th percentile (P99), which shows significant delays, Qwen 3.5-9B recorded 1.036 seconds, significantly faster than Opus 4.5's 3.312 seconds. P99 indicates that 99% of all requests were processed within that time, meaning that even in situations where the service is congested or responses are delayed, the processing speed of the Qwen model remained relatively stable.

We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it outperformed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).

I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models outperform on more narrow tasks, either working on teams, or by themselves.

We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:

"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"

I imagine we'll have some backend services that could be replaced entirely by this new architecture.

--- Brian Armstrong (@brian_armstrong) October 11, 2026

Coinbase recently revealed the process of building a payment fraud detection system for its digital asset purchase service "Onramp" in a blog series.

Onramp is a service that allows users to directly purchase digital assets from partner applications. Users without a Coinbase account can also make payments as guests, necessitating the detection of fraudulent transactions even in the absence of past account information.

Coinbase added a large language model (LLM)-based risk agent on top of its existing machine learning (ML) models and rule-based detection systems. While the existing system quickly reviews individual transactions, the LLM agent analyzes abnormal patterns across multiple transactions by looking at recent transaction sequences, time intervals, changes in payment methods, and transaction destinations.

In online experiments, the group that added the LLM agent to the existing ML model and rule engine saw a 30% reduction in fraudulent transaction counts and a 22% decrease in fraudulent transaction amounts compared to the existing system.

However, this system is different from "agent payments" where AI agents execute payments directly. The AI assesses the risk level of transactions without performing the payments, while actual approvals or blocks are handled by the existing application’s risk management system.

Coinbase initially used Claude Opus for the early system. It input transaction information of about 7,000 tokens and output only the risk level instead of lengthy explanations to reduce response time. The median latency for LLM calls was about 1.5 seconds.

Subsequently, Coinbase trained Qwen 3.5-9B using its own data instead of relying on external general models. According to the company, this model outperformed Claude Opus 4.5 in all four fraud detection metrics. The F1 score, indicating overall fraud detection performance, was 9.6% higher than that of the frontier model, and the response latency in a real service environment was reduced by 55%.

Coinbase is also establishing a fine-tuning environment that allows employees to create specialized models internally. CEO Armstrong explained that by entering commands in the internal AI system, models like Qwen can be trained for specific tasks and deployed in service environments.

An example command he provided was, "I want to fine-tune Qwen for payments risk review. Help me get started."

Armstrong suggested that this structure could potentially replace entire existing corporate systems beyond just being a simple AI assistant tool. He stated, "I believe some backend services could be completely replaced by this new architecture."

The Coinbase case illustrates that companies will likely move towards building specialized models trained on their own data instead of simply adopting general-purpose large models. In areas like fraud detection, where the scope of work is clear and correct data is accumulated, smaller specialized models can surpass general models in terms of cost, speed, and accuracy.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.