Coinbase says newer large language models underperformed older versions in fraud detection tests

Coinbase says newer large language models underperformed older versions in fraud detection tests

N
News Editor
2026-10-12 14:33:24
Coinbase said newer generations of several large language models performed worse than their predecessors in the company’s internal fraud-screening tests. In a post from its developer platform account, the exchange said it evaluated models using 16,140 transactions, including 813 confirmed fraudulent ones. The comparisons covered Opus 4.5 versus Opus 5, Sonnet 4.6 versus Sonnet 5, and GPT-5.4 versus GPT-5.6. According to Coinbase, the newer models trailed the older versions on nearly every metric in the test set. The company said possible reasons include shifts in decision boundaries and prompt behavior, changes in training data composition and task prioritization, and trade-offs introduced during post-training. Based on those results, Coinbase said it has moved the risk-control agent for its fiat on-ramp service, Onramp, to open-weight models and is carrying out domain-specific post-training.

Coinbase said newer versions of several large language models performed worse than older ones in the company’s fraud detection testing.

In a post from its developer platform account, Coinbase said it tested multiple models on 16,140 transactions, including 813 confirmed fraudulent transactions. The comparisons covered Opus 4.5 versus Opus 5, Sonnet 4.6 versus Sonnet 5, and GPT-5.4 versus GPT-5.6.

According to Coinbase, the three newer models lagged their predecessors on nearly every metric.

The company said possible explanations include changes in decision boundaries and prompt behavior, shifts in training data composition and task prioritization, and trade-offs introduced through post-training.

Based on the test results, Coinbase said it has shifted the risk-control agent for its fiat on-ramp service, Onramp, to open-weight models and is conducting domain-specific post-training.

This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan.
2300

Disclaimer:

The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.

Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.