Gartner bets open models will run 90% of enterprise agent steps by 2030

Open-weight models could take over 90% of enterprise operational agent workflow steps by 2030, Gartner analysts argue, according to a November 3 report by The Register. The pitch is a shift in where companies spend: reserve frontier reasoning for the difficult work and route the routine operations to cheaper models under their own control.

The Register’s Simon Sharwood attributes the forecast to Gartner senior director Mark Petty and group vice president Raymond Paquet. Their document appeared the previous week; November 3 is the date of the news report, not a newly established publication date for the underlying research. The 90% figure is a strategic planning assumption for 2030, rather than a measured share of today’s deployments.

The cost claim has two different denominators

The analysts describe operational workflows with repeated classification, tool calling, data passing, validation and reconciliation. As quoted by The Register, they say open-weight models can handle routine steps at comparable scores for a quarter of the price of the cheapest closed model, or as little as a twentieth of the cost of using a frontier model for every step.

Those headline comparisons do not translate directly into a twentyfold saving on completed tasks. The same report quotes the analysts acknowledging that current open models can generate more text and retry more often, bringing the saving per task against a frontier model closer to five or six times. These are Gartner’s reported comparisons; AITrending has not reviewed the underlying evaluation or reproduced them. They do not establish that every open model matches every frontier model on arbitrary work.

Routing becomes the product

Gartner’s broader argument places value in the software that runs and manages agents: routing each job to a suitable model, enforcing guardrails and managing the fleet. The analysts also argue that keeping routine work within an organization’s perimeter preserves control over accumulated workflow knowledge. Downloadable weights alone do not establish where a particular deployment sends its data.

A small November 3 Hacker News discussion illustrates the practical trade-off. One commenter reported good coding results with DeepSeek, while another preferred GLM for documentation and described using the models to check each other’s work. Those are individual experiences, not a controlled test or proof of Gartner’s forecast. The useful question for builders is which repeated steps can move to a cheaper model while preserving end-to-end quality, rather than whether a single model can replace all reasoning work.