A newer AI model does not necessarily detect fraud better. Coinbase published a technical series this week about its agent for reviewing Onramp transactions. Its results raise a practical question for crypto platforms: how to improve detection without confusing a strong general score with demonstrated financial security.
AI-assisted reporting.
Virtual editorial direction: Leonidas Ferraz.
English edition: Lin Wei.
Edition dated October 11, 2026.
A specific test, not a universal ranking
In its October 7 document, Coinbase reports that later versions of Opus, Sonnet and GPT achieved lower fraud recall and F1 scores in its test, even when precision improved. The company uses this finding to explain why a model upgrade must be assessed against known outcomes for the problem being addressed.
What the metrics mean: precision asks how many alerts were correct; recall asks how much actual fraud was detected. A tool can issue fewer incorrect alerts while missing more cases. F1 combines those dimensions, but does not by itself capture the monetary value of losses.
The specialized model Coinbase reports
On October 8, the company described tuning Qwen3.5-9B with its own fraud data, reinforcement learning through SkyRL and training infrastructure from Anyscale. It used Ray Serve LLM and vLLM to serve the model. According to that internal evaluation, the model outperformed Opus 4.5 on all four detection metrics examined, with F1 9.6% higher and end-to-end latency 55% lower.
These figures belong to Coinbase’s reported work. They are not an independent SagaTVEC audit and do not demonstrate that the same outcome would occur at another exchange, in a wallet or for an individual portfolio.
What this means for the crypto sector
Editorial analysis: the technological development concerns specialization and risk evaluation. A general-purpose model may be useful, but deciding whether a transaction is suspicious requires data and criteria specific to the system. The comparison also shows why announcing the adoption of AI does not establish that a platform’s customers are protected.
The series describes an additional review layer on top of existing machine-learning controls and rules. It does not argue that those controls should be removed or guarantee the elimination of fraud. This report does not recommend buying assets, moving funds or subscribing to services.
Sources and dates
- Coinbase, part 2: model evaluation, October 7, 2026.
- Coinbase, part 3: specialization and deployment, October 8, 2026.
Review dated October 11. Technical coverage of this week’s material, not an announcement made today. Results are attributed to the provider and were not replicated by this newsroom.
