Open models can do 90 percent of agentic work at a twentieth of the cost of a frontier LLM: Gartner

For-profit companies creating models are focusing on just a tenth of the steps in AI workflows that involve agents, as revealed by Gartner. In their report, senior analyst Mark Petty and group vice president Raymond Paquet state that nine out of ten steps are routine and can be managed by open-weight models at the same level as the cheapest closed models, but at a fraction of the cost.

These routine tasks include classification, tool calling, passing, validating and reconciling data. Open models can handle hundreds of these routine steps for every one that requires advanced reasoning. At the same performance level as the cheapest closed model, open models cost a quarter as much, and compared to a frontier model running every step, they are as little as a twentieth of the expense.

The analysts acknowledge that open models may be more verbose and retry more often, resulting in a saving of around five or six times per task. However, they emphasize that the primary benefit of open models is preserving control of data and operations. Gartner predicts that by 2030, 90% of enterprise operational agent workflow steps will be executed using open-weight models.

Enterprises will predominantly adopt domain-specific open-weight models for highly stable, recurring steps. Harnesses, which route tasks to the most suitable model, manage agent fleets, and impose strong guardrails, are the key to cost savings and operational efficiency. Frontier models will continue to play a role in research and development, serving as a capital expense, while open-weight models become an operational expense after the heavy lifting has been done.