Enterprise agents with open-weight models: Gartner's cost report, the math and the 90% rule

Gartner just published a report that, read by the headline alone, looks like one of those eye-catching predictions. But inside there are three numbers that change the bill of an agent that actually runs in a business: 9 of every 10 steps in an agentic workflow are routine, and by 2030, 90% of those steps will run on open-weight models. Why? Because open-weight models handle those repetitive steps at a twentieth of the cost of a frontier model — and in the real world, with the reformulations and retries, they still stay 5 to 6 times cheaper (the two analysts, Mark Petty and Raymond Paquet, put it in writing).
The practical question is a single one: what does this change for you if you are setting up an agent that works every day? The answer is that the way you spend your budget changes: it is no longer "which model do I buy", but "which model do I use for which step".
What an agentic step is, and why 9 out of 10 are identical
An agent does not make one big move. It breaks a goal into many small steps: classify a document, call a tool, validate a datum, reconcile two tables, generate a prompt, recheck the output. In a typical operational workflow, the two analysts estimate that 9 of every 10 of these steps are routine — classification, tool calling, data validation, reconciliation. These are steps repeated thousands of times a month, with the same schema.
The frontier model, the big and expensive one, excels at the steps that remain: the 10% that demand real reasoning, creativity, the "thinking" that does not repeat. There, it is worth paying. But pushing 9 routine steps out of 10 through a frontier model is like using a truck to carry groceries home. It works, certainly — but burning fuel and paying per kilometer for bicycle work is a waste that shows up on the invoice.
The three numbers, in a row
The numbers from the Gartner report, as picked up by The Register, are these:
| Scenario | Cost of a routine step |
|---|---|
| Open-weight vs the cheapest closed model | parity in quality, at 1/4 the price |
| Open-weight vs a frontier on every step (theoretical) | 1/20 |
| Open-weight vs a frontier (real, with reformulations and retries) | 1/5 – 1/6 |
The jump between the second and the third row is the detail that whoever sells "AI is free" forgets. Open-weight models today generate longer responses and need more attempts: so the real saving against a frontier is 5 to 6 times, not 20. But 5 to 6 times, on 9 steps out of 10, repeated every day, is a figure that changes the bill.
Let's do the math with a number
Take a concrete case, with illustrative values to get the order of magnitude (real API prices vary: the point is the structure, not the exact number).
A back-office agent makes 1,000 steps a day, 30 days a month: 30,000 steps/month. Following the Gartner data, 9 out of 10 are routine, so 27,000 routine steps and 3,000 "hard" steps.
- The 3,000 hard steps go to the frontier model, via API. A step of about 2,000 tokens (1,000 input, 1,000 output), at an indicative price of €0.005 per 1,000 tokens in input and €0.025 per 1,000 in output, costs about €0.03 per step. Total: 3,000 × €0.03 = €90/month.
- The 27,000 routine steps run on your hardware, on an open-weight model. The workstation that serves them (a consumer GPU, say 400 W under load) runs about 300 hours a month: 0.4 kW × 300 h = 120 kWh. At €0.25/kWh that is €30/month of electricity, plus the amortization of the machine: a €2,500 station divided over 36 months is about €69/month. Local total: ~€100/month for all 27,000 routine steps, i.e. about €0.0037 per step.
The sum of the mix: €90 + €100 = €190/month. If instead you sent all 30,000 steps to the frontier, you would do 30,000 × €0.03 = €900/month. The open-weight + frontier mix saves you ~€710/month, about 80% on the part that counts. And on the routine part alone, the comparison is €810 (all API) against ~€100 (all local): a ratio of about 8 to 1, the same order as the 5 to 6 times that Gartner measures.
The point is not the 710 or the 712. It is that the numbers split into two parts that are paid differently: the hard part is paid per step (API, variable), the easy part is paid once (hardware, fixed).
The part the report does not say out loud
There is a passage the two analysts put at the center, and it is worth more to a business than the saving: "Open-weights preserve your control. The cost saving is a bonus." Every time you send a routine step to a frontier API, you move a slice of your workflow's history outside your perimeter. Open-weights keep the execution inside. Control over the data is the main advantage; the money saved is the consequence.
This is exactly the reason local AI exists: not to save two cents, but to decide yourself where the work runs.
Where the value actually is
The report also moves the notion of "where the value is". It is no longer in the most expensive model: it is in the harness, the orchestration layer that decides which model to use for which step, without you having to think about it. The businesses that treat most models as commodities and invest in the router — "routine step → open-weight, hard step → frontier" — are the ones that capture the advantage. In accounting terms, the same idea: the frontier is a capex (research, paid when it is needed), the open-weight is an opex (operations, running continuously at almost fixed cost).
For you, in practice, it means two things:
- Do not choose a single model. Choose two: a local one for the repetitive work, a frontier one for the hard reasoning.
- Spend well on the hardware. A workstation that serves the 27,000 steps a month amortizes so that you no longer have a price per step: it is a fixed cost that dilutes. If you want to understand how to size it, look at How to choose your first local AI model and How to read a model's specifications.
For the general math on what a token costs, the starting point is How much does a token cost: five paths, the math and the break-even. And for the local-vs-API comparison done with the numbers, Costs: local vs API, done with the numbers.
The limits, honestly
The report is a 2030 forecast, not a present-day figure. The 90% of steps on open-weight is where businesses will go, not where they are. And the 5-to-6-times ratio is Gartner's estimate for the average mix: your real steps, your quality requirement, your hardware can shift the number in both directions. If your "hard" steps are more than 10%, the saving shrinks. If your hardware is slow and the retries increase, the local part degrades too. The numbers are a structure to reason with, not a promise: put your own values in the same formula and see what comes out.
Sources
- Open models can do 90% of agentic work at a twentieth of the cost of a frontier LLM: Gartner — The Register
- Gartner predicts open-weight models will run 90 percent of agentic steps by 2030 — summary of the report
The API prices used in the example are illustrative values to show the structure of the calculation: check the current prices of the model you intend to use.
In summary
Gartner estimates that 9 agentic steps out of 10 are routine and that by 2030, 90% of them will run on open-weight models. The reasons: open-weight models handle repetitive work at a quarter of the price of the cheapest closed model, at a twentieth of a frontier (theoretical) and at 5 to 6 times (in reality). For an agent that works every day, this changes the cost structure: the hard part is paid per step via API, the easy part is paid once, as hardware. The main advantage is not the saving — it is control, of which the saving is the bonus. The concrete move: two models, a router that decides per step, and hardware sized to absorb the repetitive work at fixed cost.
Article generated with AI
In these reading paths
Enjoyed this article?
Subscribe to the newsletter for new tutorials and guides.

