Caudena published a benchmark of 11 model configurations running agent workflows on its Prism investigation tools, scored on graph recall, graph precision and required report findings across Bitcoin, Ethereum and TRON tracing cases, with a 25 minute and $3 cap on each attempt. Across 69 scored attempts run 25 to 28 September, Gemini 3.8 Flash at medium reasoning effort scored 100 at an average of 3 minutes 40 seconds and about $0.32 per investigation, tied on quality with Claude Opus 5.5 and GLM-5.3.
Our readInvestigation and compliance teams get a stated method, time and cost figures, and the named models behind Caudena's agents, which feeds model validation. Caudena notes there is no comparison with human investigators and that a perfect score does not prove every narrative claim was supported.