Coinbase Fine-Tuned AI Model Cut Fraud Detection Costs
The company reduced latency and fraud losses by training a smaller, specialized model for its Onramp service.
Updated on Oct. 11, 2026 in Artificial Intelligence
Live Poll
Do you trust specialized, smaller AI models to perform tasks better than general-purpose frontier models?
Coinbase has successfully fine-tuned the Qwen3.5-9B model to act as a fraud detection agent for its Onramp service. The new system demonstrated improved performance over larger frontier models while maintaining low operational requirements.
Why it matters
This development highlights a shift toward using smaller, fine-tuned models for high-frequency financial tasks to gain efficiency and reduce operational costs. It demonstrates that domain-specific agents can outperform general-purpose models in specialized risk assessment benchmarks.
The fine-tuned Qwen model achieved a median latency of 0.683 seconds, which is a 55% reduction compared to the 1.515-second latency of the Opus 4.5 model. Training the model cost less than $100.
The players
Coinbase
A major cryptocurrency exchange platform that provides services including the Onramp fiat-to-crypto gateway.
NVIDIA
A leading designer of graphics processing units and specialized hardware for AI model training and deployment.
The details
Coinbase trained the model using reinforcement learning—a method where an algorithm learns to make decisions by receiving rewards for correct actions—on a single NVIDIA RTX PRO 6000 96GB GPU. The agent operates as an additional layer atop existing machine-learning models, providing a risk level assessment for individual transactions within a benchmark set of 16,140 entries including 813 confirmed fraud cases.
Timeline
October 7-8, 2026: Coinbase published the results as part of its Owning Intelligence blog series.
The Tech Race
The project follows a growing trend of enterprises abandoning massive, general-purpose models in favor of smaller, specialized architectures for specific service workflows. It marks a departure from reliance on frontier-grade models like Opus 4.5 for tasks where latency and cost are critical performance metrics.
The implementation aims to secure user transactions on the Onramp service by detecting and preventing fraud more rapidly. Users benefit from these efficiency gains through increased transaction security without needing to adjust any personal settings or workflows.
The takeaway
The success of this model suggests that high-performance fraud detection does not require the most expensive general-purpose compute resources. Watch for further benchmarks from the company to see if these improvements hold as the volume of total transactions on the platform scales.
Further reading
For more on how companies are deploying specialized agents for financial security, visit the Artificial Intelligence section.
Source note: This article includes information reported by Crypto Briefing.
Live Poll
Do you trust specialized, smaller AI models to perform tasks better than general-purpose frontier models?
