Open models can do 90 percent of agentic work at a twentieth of the cost

Enterprise AI agents are expected to perform most of their work through repetitive steps - classifying information, calling tools, passing and validating data, and reconciling results - rather than difficult reasoning. Gartner analysts argue that open-weight models can handle roughly 90 percent of these routine operations at comparable quality. Their estimates put open models at one-quarter the cost of the cheapest closed model with a similar score, and as little as one-twentieth the cost of using a frontier model for every step. In practice, greater verbosity and repeated attempts reduce the typical saving against frontier models to about five or six times. The analysts’ planning assumption is that open-weight models will run 90 percent of enterprise operational agent steps by 2030, particularly in stable, domain-specific tasks repeated at high volume.

The case for open models is not only lower expense: keeping routine work on models an organization can control helps keep operational data and workflow knowledge within its boundaries. Model capability itself is becoming a commodity, while the more valuable component is the “harness” that coordinates agents, selects suitable models, applies guardrails, and manages fleets without burdening users. Strong harnesses can route routine jobs to cheaper open models and reserve frontier models for tasks that require advanced reasoning. In this view, frontier models remain important for research and development, while open-weight models become the lower-cost operational engines built on that work.