Coinbase post-trains an open-weight 9B model that beats Opus 4.5 at fraud detection

Coinbase says it post-trained Alibaba's open-weight Qwen3.5-9B on its own Onramp fraud data and the result outperformed Anthropic's Claude Opus 4.5 on all four detection metrics in its benchmark: +7.5 percentage points on precision, +12.0 on recall, +9.6 on F1 and +35.4 on dollar-weighted recall.

How it was built
The team used reinforcement learning with verifiable rewards (scored against known historical fraud outcomes), GRPO and LoRA via SkyRL on Anyscale. Training ran on a single NVIDIA RTX PRO 6000 96GB GPU for under $100. Served with Ray Serve LLM and vLLM, median end-to-end latency on live traffic was 0.683s versus 1.515s for Opus 4.5 (55% lower), with P99 at 1.036s versus 3.312s.

Why it matters
It is the final part of Coinbase's three-part "Owning Intelligence" engineering series; part two found that newer Opus, Sonnet and GPT releases scored lower on recall and F1 for this task. CEO Brian Armstrong wrote on X that frontier models are great generalists but specialist models will increasingly win on narrow tasks, that Coinbase is making internal fine-tuning easy for employees, and that some backend services could be replaced entirely by this architecture.

Sources
Coinbase blog: Part 3/3 Owning Intelligence Series (Oct 8, 2026); Brian Armstrong on X