Open models can handle 90% of agentic workflow steps, Gartner says, at a fraction of the cost

The claim landed on Tuesday morning. In a document titled "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps," Gartner senior director Mark Petty and group vice president Raymond Paquet argue that most agent work is routine. Classification, tool calling, passing, validating and reconciling data make up the bulk. Open-weight models, they write, "can handle hundreds of these routine steps for every one that needs frontier reasoning." The Register reported the document on 3 November.

"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.

Then comes the caveat, which matters more than the headline number. The analysts admit 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." That is still a large gap. But it is a different order of magnitude from the one in the press coverage.

Where the money and the control both go

Gartner's argument is not purely about price. The analysts frame open weights as a data control question first. "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. Open-weights preserve your control. The cost saving is a bonus." Their formal planning assumption is that by 2030, 90% of enterprise operational agent workflow steps will execute on open-weight models.

That projection sits awkwardly against the frontier releases of the past week. Tom's Hardware reported on 3 November that Google's Gemini 4 Carbon (high) scores 59 on Artificial Analysis' Intelligence Index V5. That is one point ahead of Claude Opus 5.5 (max with fallback) and two ahead of OpenAI's GPT-6 Astra (max). The same article puts Gemini 4 Carbon's cost per task at $7.76, against $12.99 for Claude Fable 5.1. That is roughly 13 times the per-task price of the open-weight DeepSeek V4.1 Flash 1017, which scores 32 on the same index.

Gartner's answer to that gap is that the frontier model becomes capital expense and the open model becomes operating expense. "Frontier models will continue to have a role as the engine of research and development," The Register summarises the paper as saying. Open-weight models take over once the heavy lifting is done. The analysts also point at harnesses, the software that routes each task to a suitable model, as the place where product value ends up.

The research stack is moving the same way

Two arXiv papers submitted on 29 and 30 October suggest the training side is being reorganised along similar lines. QUAL, a framework from Junhyeok Kang and 11 co-authors, treats a pool of time-series foundation models as a probability reference. It transforms each member's forecasts before combining them. It reports reductions of 7.4% in average member MASE and 9.8% in CRPS on GIFT-Eval. It requires no training and no model modification.

Separately, Yuda Song and nine co-authors argue in a 67-page paper that pretraining, midtraining, supervised fine-tuning and reinforcement learning can sit under a single likelihood-based algorithm. That collapses the staged pipeline into a question of which data to train on and when.

Neither paper tests Gartner's cost thesis. Both point at the same direction of travel, though: fewer bespoke stages, more reuse of whatever models already exist.

The hardware side is making the same bet. The Robot Report reported on 2 November that Tutor Intelligence raised $75 million in Series B funding led by Andreessen Horowitz. Customers pay by the hour or by the pick rather than buying robots outright. Agility Robotics, meanwhile, cleared its SPAC vote on 2 November with more than 84% of votes cast in favour. The combined company is expected to start trading on Nasdaq on 5 November under the ticker AGLT, with roughly $465 million in gross capital before expenses. Neither number settles whether routine work migrates to cheap models. Both assume somebody is counting the cost per task.