Revolut moved 2,000 developers to GLM and cut AI costs by 90%

When AI coding assistants first appeared, their bills looked trivial compared with salaries. Then developers began running agents for many hours a day, and that trivial expense grew into a separate budget line. At Revolut, it reportedly reached as much as $10,000 per month for some engineers. The fintech company decided the numbers no longer added up.

What exactly they did

According to reports, Revolut moved around 2,000 developers from Claude Code and Codex to the open GLM model from China's Z.ai. One million tokens now cost about $0.30, which is 10 times cheaper than before. The company's total AI spending reportedly fell by 90%.

The key detail here is not the price but what happened to the limits. Engineers previously kept a mental counter running: whether to send the entire repository to the model, run the agent one more time or ask for a second solution. When tokens are expensive, people start rationing questions, and that hurts work quality more than it might seem. Revolut removed token limits entirely.

What the numbers show

According to the publication, code-writing time fell by 60%, while task completion time dropped by 70%. The logic is straightforward. If an agent can freely read context, double-check its work and rewrite code, it makes fewer foolish mistakes, while the human spends less time waiting and intervening.

Consider a simple example. An engineer needs to migrate a service to a new library version. With a strict limit, they feed the model two or three files and edit the rest by hand. Without a limit, they give the agent the entire module, including its tests, and ask it to keep working until everything is green. This consumes many times more tokens, but at $0.30 per million, that is no longer a problem.

Where caution is warranted

First, the figures currently come from a retelling and a single publication. There has been no independent audit, and we do not know exactly how Revolut calculated the savings, how many tasks remained on closed models or what share of the edits engineers still make manually. A headline saying they were "completely replaced" can easily be interpreted more broadly than the facts support.

Second, a 90% saving on tokens does not equal a 90% overall gain. If GLM performs worse on the most complex tasks, some of the saved time will be spent on revisions. The 60% and 70% metrics look impressive, but it is unclear which tasks were measured or what comparison was used.

Third, for a fintech company, price is not the only issue. Open weights make it possible to keep the model inside the company's own environment, which is an important consideration for customer data. But that adds infrastructure, monitoring and the team maintaining it to the bill. Using the model through Z.ai's API raises a different set of questions around jurisdiction and compliance.

What to do in practice

Start by calculating how much AI coding actually costs per engineer in your organization. If the figure is approaching thousands of dollars per month, a comparison with open models is already justified.

Next, run an honest pilot. Take real tasks from your team, run them through both a closed and an open model, and compare not only the price but also the acceptance rate of the proposed changes. Test the effect of removing limits separately: it may be the main source of the speed increase rather than the specific model.

One final point. This story shows that the coding-assistant market is ceasing to be a one-leader contest. When an open model handles a large share of routine tasks at one-tenth of the price, closed providers will have to prove their advantage on the hardest cases, not on average code.