[Coinbase: New Generation of Large Models Perform Worse Than Older Versions in Fraud Detection Tests] Foresight News reports that the Coinbase Developer Platform account tweeted that it conducted risk control tests on several large models using 16,140 transactions (813 of which were confirmed fraudulent). The tests compared Opus 4.5 with Opus 5, Sonnet 4.6 with Sonnet 5, and GPT-5.4 with GPT-5.6. The results showed that the newer versions of all three models performed worse than the older versions across almost all metrics. Coinbase believes the reasons may include changes in decision boundaries and prompt behavior, changes in the composition of training data and task priorities, as well as trade-offs introduced by post-training. Based on the test results, Coinbase has shifted its Onramp risk control Agent to open-weight models and conducted post-training for specific domains.