Coinbase's 9B Fraud Model Beat Opus 4.5: Copy the Pattern With a Verifiable Reward, Tier-Only Output, and Block-Only Harness
Attributed compile + one small offline run on synthetic data Primary sources (Coinbase Engineering, "Owning Intelligence" series): Part 1/3 — Build by Yao Ma, Bo Lei, Xuwei Tan (2026-10-06), Part 2/3 — Evaluate by Yao Ma, Xuwei Tan (2026-10-07), and Part 3/3 — Own by Yao Ma, Xuwei Tan, Akshit Trehan, Aman Choudhary with the Anyscale team ( Eric Tang, Sumanth Hegde, Seiji Eicher, Jess Kong )…
Coinbase's 9B fraud model outperformed Claude Opus 4.5 on Coinbase's own fraud benchmark. The model, trained on the company's fraud data, was developed in three parts. First, a layered system where an existing ML model and rules engine screen transactions before an LLM agent reviews selected transactions. The agent outputs a risk level without explanation.
Second, a block-only authority where the agent can only block transactions it allowed upstream. Performance was tested against Opus 4.5, with the Coinbase model achieving higher precision, recall, and dollar-weighted recall. The model was served using Ray Serve on L40S GPUs, with P50 latency of 0.683 seconds compared to Opus's 1.515 seconds.
The reward function is deterministic, based on the model's output and known label, and dollar-weighted recall is used instead of F1 to account for the value of fraud transactions.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
