Two Gartner analysts argue cheaper open models will handle most routine AI agent work, with value shifting to orchestration
Gartner thinks most of the work done by enterprise AI agents will soon run on models that companies can download, inspect, and run themselves.
Analysts Mark Petty and Raymond Paquet forecast that open-weight models will power 90% of enterprise agentic workflow steps by 2030. Their reasoning is mostly about money, though control comes up too.
The math behind the forecast
The two analysts focus on the unglamorous parts of an AI agent’s job. That means classification, tool calling, data validation, and reconciliation.
According to Petty and Paquet, open-weight models can now handle these steps at quality comparable to closed models. Once quality is matched, the price gap becomes hard to ignore.
“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,” the pair wrote.
Put plainly: if two models score the same on a task, the open one can cost 25% of the cheapest closed rival. Compared with sending every step to a top-tier frontier model, it may cost one-twentieth as much.
The analysts temper that headline figure with a more grounded estimate. Once redundancy and operational inefficiencies are factored in, they put real-world savings at roughly 5 to 6 times.
Why volume changes everything
Petty and Paquet expect enterprises to favor domain-specific, fine-tuned open-weight models for stable, high-frequency tasks.
The analysts describe these as jobs executed millions of times per month. At that scale, small per-step savings compound into budget line items.
An open-weight model is one whose trained parameters, the numerical guts of the model, are published. Companies can host them on their own infrastructure, modify them, and avoid paying per call to an outside provider.
Closed models work differently. You rent access through an API, the provider sets the price, and the model’s internals stay behind the curtain.
The analysts also make a point that goes beyond cost. “Open-weights preserve your control,” they note, a reminder that self-hosting means a company is not at the mercy of a vendor’s pricing changes or product decisions.
The value moves to the conductor
Petty and Paquet caution that value will migrate away from the raw capabilities of individual models, which they see as becoming commoditized.
Instead, the value shifts toward orchestration. That is the layer that ties multiple agent workflows together and decides which model handles which step.
What this means for vendors and buyers
For enterprise buyers, the analysis offers a practical playbook. Sort agent tasks by difficulty and frequency. Send stable, high-volume work to fine-tuned open models. Reserve expensive frontier calls for the steps that actually need them.
The catch is that self-hosting is not free. Running open models means owning infrastructure, monitoring, and maintenance. That is likely part of why the analysts’ practical savings estimate sits at 5 to 6 times rather than the full twentyfold figure.
There is also a strategic dimension. A company that relies entirely on one closed provider has handed over a lot of leverage. Spreading routine work across open models reduces that dependency, which is the control point Petty and Paquet keep returning to.
