Coinbase disclosed on October 7 that newer iterations of three prominent AI model families identified fewer fraudulent payments and captured a smaller portion of fraudulent value during a historical evaluation of its Onramp payment screening system, even though the fraud detection policy remained unchanged. The results question the assumption that advancing a model automatically enhances fraud detection performance. The company replayed 16,140 transactions from 7,293 users, including 813 confirmed fraudulent cases, drawn from nine weeks of activity prior to the deployment of its risk agent. This cohort preserved all documented fraud incidents while sampling legitimate transactions. Each candidate model assessed recent transaction behavior under fixed guidance and the same decision policy, isolating the model's behavior within that framework rather than comparing redesigned screening systems. Coinbase compared Opus 4.5 with Opus 5, Sonnet 4.6 with Sonnet 5, and GPT-5.4 with GPT-5.6 (sol). Every newer version recorded lower recall, reduced combined precision-and-recall scores (F1), and diminished dollar-weighted recall. Recall reflects the proportion of fraud cases detected; dollar-weighted recall measures the value of fraud caught. Sonnet's recall declined by 22.2 percentage points and its dollar-weighted recall dropped 22.9 points. Opus showed only a 0.8-point recall decrease. Both newer models also exhibited lower precision, meaning fewer of the transactions they flagged as fraudulent were actually fraudulent. GPT's performance illustrated how one improving metric can be misleading: its precision increased by 11.5 percentage points, yet recall fell 20.7 points and dollar-weighted recall dropped 21.8 points, resulting in more fraud cases and value escaping detection. The replay does not quantify customer losses from deploying these versions. Coinbase noted it could detect the performance regressions without determining their root cause. Earlier experiments had compared incorporating selective large language model review with existing models and rules alone, which recorded 30% fewer fraudulent transactions and 22% less fraud value, though that test did not evaluate newer model versions. In a separate disclosure on October 8, Coinbase reported that a post-trained Qwen3.5-9B model outperformed Opus 4.5 across four fraud detection metrics, improving F1 by 9.6 percentage points and increasing dollar-weighted recall by 35.4 points. This model was fine-tuned using historical fraud outcomes and balanced rewards for fraudulent and legitimate examples. Production measurements indicated a median end-to-end latency of 0.683 seconds for the Qwen3.5-9B model versus 1.515 seconds for Opus 4.5, representing a 55% relative reduction. Coinbase emphasized that for payment providers, the critical question is whether a candidate model improves fraud coverage under their actual decision setup, recommending configuration testing before adjusting prompts or thresholds, while also considering latency, reliability, and cost alongside detection quality.