Open models may lower agent costs

Gartner’s claim, reported by The Register, is that open-weight models can handle 90% of enterprise agent workflow steps. The striking number is cost: against a frontier model used for every step, open models could cost as little as one-twentieth as much. But that’s a best-case comparison, not the likely saving on a whole task. Open models can be more verbose and retry more often, bringing the saving per task closer to five or six times.

The business idea is to send routine work, such as classification, tool calls and data checks, to a cheaper model, while reserving frontier models for steps that need more reasoning. Gartner argues that the software coordinating this work, called a harness, matters because it routes each step to a suitable model and manages guardrails. That makes model choice only part of the decision. A company also needs a way to direct work reliably.

For an operator, the useful question isn’t whether every task should use an open model. It’s which steps are stable and routine enough to test with one, and whether the cost saving holds once retries are counted. This is the kind of model routing and oversight we build into an AI command centre.