Today · Tuesday, October 13, 2026
Enterprise AICoinbaseOctober 8, 20262 min
Coinbase post-trained a 9B model that beat Opus 4.5 on its fraud test
Company claim
Reported by Coinbase; the benchmark is the company's and is not independently verified.
Coinbase · October 8, 2026 · Primary
What happened
On October 8, 2026, Coinbase published the third part of its Owning Intelligence series. The company post-trained Qwen3.5-9B, an open-weight model, for fraud detection on Onramp, using SkyRL on Anyscale, and served it with Ray Serve LLM and vLLM. Coinbase says that on its benchmark the tuned model beat Claude Opus 4.5 on all four detection metrics, including a 9.6 percent higher F1 score, and cut end-to-end latency by 55 percent. The day before, the same series said newer Opus, Sonnet, and GPT versions scored lower on recall, F1, and dollar-weighted recall for this task than the versions they replaced.
Why it matters
Frontier models are generalists. This is a company taking a small open model, training it on data only that company has, and using the result for one decision. Coinbase says it is making that fine-tune path available to its own employees, which is the part other companies can copy without training a foundation model. The scores are Coinbase's. The benchmark is not public, so this is an internal measurement, not a leaderboard anyone else can rerun.
Go deeper
The series is three posts, October 6 to October 8. Part 1 describes an agent layered onto Onramp's existing models and rules. Part 2 is the evaluation that pushed the team off the newest frontier releases. Part 3 is the post-training.
Coinbase's chief executive said on October 11 that frontier models remain strong generalists, and that specialist models will win more narrow tasks. He said the company is making internal fine-tuning easy for employees. That claim is his. The measurement behind it is the October 8 post.
Sources
Reported by Coinbase; the benchmark is the company's and is not independently verified.
- Coinbase Primary · October 8, 2026