Inference tiering is a model-routing strategy that sends simple AI tasks to cheaper, smaller models and reserves stronger models for harder reasoning. In practice, an orchestrator may use a lightweight model for classification, retrieval, summarization, or validation, then escalate only complex or ambiguous steps to a frontier model. This reduces compute spend and latency without giving up quality where it matters.
In cyber security, inference tiering matters because agentic systems can be steered through many steps, and every step is both a cost and a security decision. Defenders use it to cap token burn, shrink the attack surface of tool-using agents, and keep routine security workflows efficient. Attackers can also abuse poorly designed routing by flooding systems with edge cases that force expensive escalations, or by probing which tasks trigger stronger models and stricter checks. Good tiering pairs model selection with permissions, logging, and stop conditions so that cost control and risk control work together.