Coinbase Reports Gains From a Post-Trained Onramp Fraud Model

Coinbase says post-training Qwen3.5-9B improved results on its proprietary Onramp fraud benchmark. The narrow question is transaction risk: how well a specialized model identifies fraud in Coinbase’s evaluation, and how quickly it responds when served on production traffic.

The findings come from three company articles published October 6–8. Reading them together requires keeping three pieces of evidence separate: an earlier experiment with an LLM risk agent, a historical benchmark, and the subsequent post-training and latency results. Those distinctions determine what the reported gains can establish.

An earlier experiment with the existing stack

In its October 6 report, Coinbase says it added an LLM risk agent after its existing Onramp machine-learning model and rules. The agent-enabled group received 50% of traffic. Coinbase reports 30% fewer fraudulent transactions and 22% less fraud value than with the existing stack alone.

That article does not disclose the experiment’s duration or sample size. These are results for the earlier agent-enabled stack, not the later post-trained model’s benchmark scores. Combining them would obscure which system was evaluated and under what conditions.

The historical benchmark has disclosed dimensions

Coinbase’s October 7 report describes a historical benchmark containing 16,140 transactions from 7,293 users across nine weeks of production data, including 813 confirmed fraudulent transactions. The benchmark’s size and time window are disclosed; they should not be confused with the earlier experiment’s missing details.

Coinbase also reports that newer Opus, Sonnet and GPT versions scored lower than their paired predecessors on recall, F1 and dollar-weighted recall in this evaluation. That is a task-specific finding. It does not establish a general ranking of model families. The company provides neither independent validation nor confidence intervals.

Post-training gains and latency are separate results

In its October 8 report, Coinbase says it post-trained Qwen3.5-9B on proprietary Onramp fraud data and served it using Ray Serve LLM and vLLM. Against Opus 4.5 on its benchmark, Coinbase reports a 9.6-percentage-point F1 gain and a 12.0-percentage-point recall gain. These are percentage-point differences, not relative percentage improvements.

Separately, Coinbase reports median, or P50, LLM-request latency on online production traffic of 0.683 seconds, compared with 1.515 seconds for Opus 4.5. Detection quality and latency were measured separately. The faster response figure therefore should not be presented as another fraud-detection score.

What readers can take from the reports

The editorial implication is that evaluation context matters: a specialized task can produce comparisons that a broad model ranking would miss. But a private benchmark cannot establish that the same gains would carry over to another platform’s transactions. Independent reproduction, uncertainty estimates and information about customer false-positive burden would make the results easier to assess.