Gartner Says Open Models Will Run 90% of Agent Steps; Revolut Says Self-Hosted GLM Cut Costs 90%: Claimed vs Verified

Gartner says open-weight models will run most of the work in enterprise AI agents, and at a fraction of the cost. In a paper titled "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps," senior director Mark Petty and group vice president Raymond Paquet argue that nine in ten steps in a typical agent workflow are routine, and that open models handle them at parity for roughly a quarter of the price of the cheapest closed model. The Register reported the paper on November 3, 2026. The same day, a separate story claimed that Revolut cut its AI coding token costs by 90% by moving about 2,000 engineers to a self-hosted GLM model. This post lays out what each source says, what is confirmed, and what is still only a claim.

TL;DR

What Gartner actually wrote

According to The Register, the paper starts from a simple observation about how agents work. A typical operational workflow involves hundreds of routine executions: classification, tool calling, parsing, validating and reconciling data. The analysts say open-weight models "can handle hundreds of these routine steps for every one that needs frontier reasoning." In their words, "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."

Those are the best-case figures, and the analysts say so. They admit current open models "can be more verbose and retry more often," which pulls the saving per task against a frontier model down to about five or six times. That is still large, but it is a different headline from "a twentieth of the cost."

The analysts also frame cost as a side benefit. Their main argument is control: years of workflow history are a large part of what an acquirer pays for in a software company, and every routine step sent to a frontier API moves a slice of that history outside the company's perimeter. Open weights keep it in house. They conclude with a forecast: by 2030, 90% of enterprise operational agent workflow steps will run on open-weight models.

There is a second claim in the paper that matters as much as the first. The analysts argue that models are becoming commodities and that the valuable part of an AI product is the harness, the software that runs agents, routes each step to the right model, applies guardrails, and manages fleets of agents. A good harness saves money without users noticing. A weak one leaves more work to the user. Frontier models, they say, remain the engine of research and development, which they treat as capital expense, while open models become the operating expense once the heavy lifting is done.

The Revolut claim

A small house-shaped box with a glowing core, standing for a self-hosted model

The Revolut story rests on a summary of a conference talk. As relayed by a StartupHub write-up (we could not open the page ourselves, so we rely on the search excerpt and an aggregator summary), David Abitbol, named in the write-up as the speaker, said Revolut moved about 2,000 engineers off metered APIs from OpenAI and Anthropic and onto a self-hosted GLM model for coding. The reported details:

• The system runs on 56 Nvidia B200 GPUs.
• The blended price is about 30 cents per million tokens, which is roughly 10% of what Revolut previously paid.
• Claude Code and Codex became exceptions rather than defaults.
• Engineers stopped rationing tokens and could use AI on all tasks, not just the hardest tickets.
• Lead time improved by 70%, and the write-up reports roughly 60% faster coding time. The write-up says the larger lead-time gain comes partly from non-coding tasks such as inbox triage and meetings.
• Caveats in the write-up: portability is unproven, the numbers assume steady tuning and a bank that can set sandbox rules and work hours to shape load, and image tasks go to other models.

The headline says "GLM." The model line we could confirm in the excerpt names GLM 4.5, but other GLM versions have shipped since, so treat the exact version as unconfirmed until Revolut or the talk itself says it.

Claimed vs verified

Verified: The Register published the Gartner summary on November 3, 2026, and quotes the paper directly. The existence of the Revolut claim is also verified: a trade write-up reports it and attributes it to Abitbol.

Not verified:

• Gartner's cost ratios. We saw no data set, task list, or benchmark names. "At the same score" depends on which score. A routine-step benchmark where small models match large ones is very different from a messy real workflow.
• The 90% share. This is a forecast to 2030, not a measurement of what enterprises run today.
• Revolut's 90% saving. The 30 cents per million tokens figure is a blended internal number. It does not obviously include GPU purchase or lease, power, the engineers who run the cluster, or the cost of the tasks that still go to frontier APIs. Whether the 56 B200s are owned or rented is not stated in what we saw.
• The 70% lead-time gain. Lead time moves for many reasons. Nothing seen separates the effect of the model from process changes, a new workflow, or the removal of token rationing.

How to check: read the Gartner document if you have access, ask Revolut for the talk recording or slides, and, most useful, run the comparison on your own workflow (see below).

Why the claim is plausible

None of this means the claims are wrong. Several independent signals point the same way, and we have covered some of them.

Coinbase said in its Owning Intelligence series that it post-trained the small open Qwen3.5-9B for under $100 and that it beat Claude Opus 4.5 on one fraud task. That result is also company-reported, and we covered the method and caveats in our Coinbase Qwen3.5-9B post. It supports the Gartner idea that narrow, stable, high-volume steps favor a specialist open model.

Independent measurement of the gap between open and closed models also exists. Our write-up of the Artificial Analysis Intelligence Index v5 shows the open-weight gap to the top closed model at about 12 points. A 12-point gap on a broad index matters for hard reasoning, and matters much less for a step like "extract the invoice number."

Large open-weight releases keep arriving. See, for example, our coverage of Atria Dawn, a 744B open-weights agentic model, and our small models comparison, which includes GLM 5.3 Flash against closed small models.

Why the claim may be too clean

A gauge with a needle rising toward its limit, standing for agent token cost per completed task

Self-hosting has failure modes that a cost-per-token comparison hides.

Operations risk. Running your own inference stack means owning its bugs. Our report on Cloudflare Workers AI and a GLM 5.3 KV cache leak shows how a serving-layer bug in an open-model deployment exposed other users' responses. A bank that hosts its own models inherits this class of bug, even if it also gains the data-control benefit.

Utilization. A GPU cluster is cheap per token only when it is busy. A cluster sized for peak load idles at night. The per-token price quoted for Revolut is a blended figure, and a team with spiky load may not see the same number.

Retries and verbosity. Gartner itself flags this. If an open model needs two attempts, or writes three times the tokens for the same answer, the per-token discount shrinks. Cost per completed task is the number that counts.

The hard 10%. If frontier models still handle the difficult tenth of steps, the real saving depends on how well the harness routes. Routing mistakes in either direction cost money: send a hard step to a weak model and you pay for failures; send easy steps to a frontier model and you pay the premium you wanted to avoid.

Safety and provenance. Open weights are not automatically safe. Weights from outside your jurisdiction raise supply-chain questions, and some labs' models have shown surprising behavior in agent settings. We covered one example in the Cohere and Qwen sandbox bucket-upload report. Agents with real tool access need monitoring whichever model drives them. For that layer, see AgentBeam, the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

What "harness" means here

The Gartner point about harnesses is worth a short explanation. A harness is the code around the model: the loop that calls tools, the memory, the permissions, the logging, and the router that picks which model handles each step. The same model can look strong or weak depending on its harness. If the analysts are right that models are commodities, the competitive question moves from "which model" to "which routing and guardrails." Our guide to an open-source agent harness with memory and skills shows what that layer looks like in practice, and the Stripe checkout token study shows a different lever on the same cost: cutting the tokens an agent has to read in the first place.

How to test this for your own workflow

1. Log a week of real agent runs. Record every step: what it did, which model ran it, tokens in and out, retries, and whether it succeeded.
2. Label each step. Routine (classify, extract, call a tool with known arguments, validate) or hard (plan, debug, write new logic, resolve ambiguity).
3. Replay the routine steps on one or two open-weight models and score success against the logged results with the same checker. Do not rely on a public benchmark alone.
4. Compute cost per completed task. Include retries, extra tokens, GPU or hosted-inference price, and the time of whoever runs it.
5. Route, then re-measure. Move only the steps that matched. Keep a frontier fallback for failures. Watch for regressions on each model upgrade, as the Coinbase replay suggested can happen.
6. Decide on control separately. If the reason is data custody, as Gartner stresses, the decision does not need the savings to be 20x.

What to watch next

Three things would turn these claims into evidence. First, a published method from Gartner or a customer that names tasks, models, and scoring. Second, a Revolut or conference release of the talk with infrastructure costs included. Third, independent replications from teams that publish their own routing results. Until then, treat "90% of steps" and "90% cheaper" as informed hypotheses from analysts and one practitioner, worth testing, not facts to plan a budget around.

Related reading

• Coinbase Qwen3.5-9B fraud agent: method and caveats
• Artificial Analysis Intelligence Index v5 and the open-weight gap
• Small models comparison, October 2026
• Cloudflare Workers AI GLM 5.3 KV cache leak
• Stripe checkout pages and AI token use
• Menlo Ventures: open models handle 1 in 5 tokens but 1 in 25 dollars