New reporting on token costs, Verizon infrastructure and Microsoft's security agents shows enterprises separating expensive frontier work from routine inference. Companies are using model routing to send simple requests to cheaper systems and reserve premium models for work that requires greater capability. Moody's reported examples in which AI token costs at large organizations doubled nearly every other month without an equivalent measured productivity increase.
Microsoft introduced MAI-Cyber-1-Flash, a compact security model integrated into a multi-agent scanning system for identifying exploitable software weaknesses. Microsoft said its system scored 96% on the CyberGYM benchmark and cost half as much as its previous offering, claims that require independent production testing.
Verizon plans to convert some network offices into smaller facilities for low-latency inference while linking Google data centers through a $1 billion dark-fiber agreement. These developments place cost, latency, task selection and review requirements alongside raw model capability in enterprise deployment decisions.
A smaller model can be faster and cheaper but may fail on difficult or unusual work, making routing accuracy and escalation rules part of system quality. Cybersecurity agents can automate high-risk actions, so benchmark performance does not replace containment, logging, access control and human review.
Edge inference reduces network delay for some applications but adds distributed infrastructure, power, maintenance and physical-security requirements. As of the edition cutoff, Vendor performance and savings claims were not independently audited across comparable production workloads. The available reports distinguish measurements and attributed claims from conclusions that had not yet been established. Where officials supplied numbers, those figures remain subject to revision as agencies reconcile records and publish later updates. Subsequent records may therefore change the totals without changing the initial chronology. The next factual records to examine are published enterprise cost and productivity measurements and independent security-agent evaluations and real utilization of new edge facilities.