Nine in ten of the steps in an enterprise agent workflow are routine, and open-weight models can handle them at the same quality for a quarter of the price of the cheapest closed model, according to a new document from the analyst firm Gartner.
The paper, "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps", was written last week by the firm's senior director Mark Petty and group vice president Raymond Paquet, and was reported by The Register.
"Nine in 10 agent workflow steps are routine, and open-weight models handle them at parity for a quarter of the price of the cheapest closed model," the pair write. They describe typical operational workflows as involving "hundreds of routine executions, such as classification, tool calling, passing, validating and reconciling data".
Open-weight models, they argue, "can handle hundreds of these routine steps for every one that needs frontier reasoning". Their strategic planning assumption is that "by 2030, 90% of enterprise operational agent workflow steps will execute on open-weight models".
The cost case is stated in two figures. "Against the cheapest closed model at the same score, the open model costs a quarter as much. Against a frontier model run on every step, it costs as little as a twentieth," they write.
They then qualify it. "Current open models can be more verbose and retry more often, so the saving per task against a frontier model is closer to five or six times," the pair add.
The analysts rate the saving as a by-product rather than the point. "Years of workflow history are a large part of what an acquirer pays for in a SaaS company and every routine step sent to a frontier API moves a slice of it outside your perimeter," they write. "Open-weights preserve your control. The cost saving is a bonus."
They expect enterprises to adopt mostly domain-specific open-weight models, which do well where "a small number of highly stable steps run millions of times a month". Models themselves, the paper suggests, are already commodities; the value sits in the harnesses that run agents and route each job to a suitable model without the user noticing. Frontier models remain the engine of research and development, effectively a capital expense, with open-weight models an operating expense afterwards.
The two cost figures in the paper are worth holding side by side. A twentieth is what the open model costs when it is compared with a frontier model. A fifth or a sixth is what it costs when it is asked to do some work.
The difference between the two is attributed, accurately, to verbosity and retries. This newspaper notes that verbosity and retries are not flaws in the comparison. They are what the model does. The twentyfold figure is therefore a measurement taken of the model before it has said anything.
Read in that direction the paper becomes more useful than it first appears. The cheapest conceivable open-weight model answers once, briefly and correctly, which is also the description of the frontier model it is being compared against and priced at a twentieth of.
The harness is offered as the remedy, and on the paper's own logic it is the right one. A harness that routes each job to the most appropriate model will send any step that genuinely needs frontier reasoning to the frontier model, which is the one step in ten where the quoted saving does not apply and the only step anybody was worried about.
The Department of Unnecessary Research accepts the 2030 assumption without amendment, and observes that a forecast in which nine steps in ten are routine, and the saving holds only on routine steps, is a forecast about the steps rather than the models, and could have been filed before either kind of model existed.