Gartner sees open models dominating AI agents

Gartner estimates that open-weight models can execute 90% of agent workflow steps at a much lower cost than closed LLMs. The challenge shifts to orchestration and control tools.

According to Gartner, open-weight models could support 90% of the steps in enterprise agent workflows, at a cost ranging from one quarter to one twentieth of that of a cutting-edge LLM depending on the case. The analyst considers that these models will mainly handle repetitive work, while closed models will remain useful for the most demanding reasoning tasks. This shift matters for companies looking to reduce their AI bill without losing control over their data and workflows.

Les tâches routinières basculent vers l’open-weight

In a note titled Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps, analysts Mark Petty and Raymond Paquet describe workflows where most actions are mundane: classification, tool invocation, transmission, validation, and data reconciliation. Their thesis is simple: nine out of ten steps do not require a state-of-the-art model, but reliable and repeatable execution. In other words, an agent is not asked to reinvent the wheel for each ticket.

Gartner states that open models can perform these steps "at parity" for a quarter of the price of the cheapest comparable closed model. In the best case, the cost can drop to one twentieth of that of a cutting-edge model used at each step. However, the analysts warn that open models can be more verbose and multiply attempts, which reduces the actual gap to five or six times in many scenarios.

Le vrai sujet : le moteur, ou l’atelier qui le pilote ?

The note emphasizes a less spectacular but more strategic point: the value is no longer only in the model itself. Gartner describes "harnesses" as the software layer that orchestrates agents, applies safeguards, allocates tasks to the right model, and manages fleets of agents. This is where part of operational control is played out, and thus part of the total cost.

The analysts believe that companies will favor open-weight models specialized in scenarios where a few stable steps repeat millions of times per month. Cutting-edge models, meanwhile, would retain a research and development role, almost like an initial investment expense, before open models take over operational spending. At this stage, Gartner does not describe a total shift, but a clearer division between what explores and what executes.

Une projection pour 2030 qui dit surtout où va le budget

Gartner puts forward a planning hypothesis: by 2030, 90% of the steps in enterprise agent workflows will be executed on open-weight models. This projection does not say that closed models will disappear. It rather suggests that they will be reserved for sequences where complex reasoning still justifies their price.

For model providers, the message is uncomfortable. If profitable tasks concentrate on a minority of steps, the commercial battle shifts to software infrastructure, orchestration tools, and security guarantees. Companies, for their part, will look less at the model's prestige than at its ability to get the job done quickly, quietly, and without sending data on unnecessary detours.

Key points

  • 90% of agent steps could shift to open-weight models.
  • Mark Petty and Raymond Paquet authored the Gartner note.
  • Potential cost of one twentieth compared to a cutting-edge model.
  • Gartner projection: 2030 for enterprise workflows.
  • "Harnesses" become the most valuable layer.
  • Tasks mentioned: classification, validation, data reconciliation.

By the numbers

  • 90% — share of agent steps targeted by Gartner, enterprises, 2030 projection.
  • 1/4 — cost of an open-weight model compared to the cheapest comparable closed model.
  • 1/20 — possible cost compared to a cutting-edge model used at each step.
  • 5 to 6 times — more realistic savings estimated when feedback is taken into account.

To read