A new arXiv preprint (2610.14180) addresses a practical dilemma for organizations building language-model agents for security: whether to use open-weight models they can run and audit internally, or pay a premium for closed frontier models. The paper frames this as a direct comparison on CTI-REALM, a benchmark apparently designed for cyber-threat-intelligence tasks.
The abstract states that the authors "study that choice on CTI-REALM," but the provided text cuts off before any findings or methodology details are given. As a result, the paper's conclusions are not available from the abstract alone.
What is clear is the motivation: open-weight models offer transparency and self-hosting, while closed frontier models typically come with higher costs but potentially stronger performance. The study aims to quantify that trade-off in a security context, though the outcome remains unknown from the source material.