Revolut Says Chinese Open-Weight AI Model Cut Its AI Costs by 90%

Revolut has shifted much of its developers’ AI workload to GLM-5.1, an open-weight model from Chinese AI lab Z.ai that the bank runs on dedicated machines. David Abitbol, who leads AI operations at Revolut, said the change lowered the company’s AI spending to around a tenth of what it had paid Anthropic, while accelerating software development.

Abitbol shared the figures during a presentation in New York in mid-October, later republished on YouTube on October 29. He said Revolut drew on data from 2,000 engineers to assess how AI tools affected coding time and delivery timelines.

The company’s approach has evolved over the past two years. In early 2025, adopting Gemini cut coding time by 20% and shortened delivery time—the period between taking on a feature and preparing it for deployment—by 10%, according to Abitbol. Revolut later expanded its use of Anthropic’s Claude Code, while imposing strict spending limits. Abitbol said some engineers were spending between 1,000 and 10,000 a month without a sufficiently strong return on investment; he did not specify a currency in the presentation as described in the source.

Revolut’s productivity figures rose in April 2026, when it moved to GLM-5.1 with no token limits for users. The bank now reports a 60% improvement in coding time and a 70% improvement in delivery time.

Unlike services commonly priced by token use, GLM’s open-weight model can be downloaded and installed on a company’s own infrastructure. Revolut says this lets its engineers use AI without worrying about individual consumption charges. The bank runs the system on 56 Nvidia B200 chips. The setup is not cost-free: Revolut pays for the machines and may need to add capacity if demand exceeds what they can handle. Abitbol put the average cost at 30 cents per million tokens and said total spending is one-tenth of the company’s previous payments to Anthropic.

Setting up the system required extensive tuning, he said. Revolut reduced the wait before a response began from more than 10 seconds to 1.6 seconds, using the same number of chips even as demand climbed from three to more than seven requests per second.

Abitbol also addressed concerns about using a Chinese-developed model. He said Revolut downloads the model’s weights and runs them on a service based in the United States, rather than using a service operated in China. He added that data must remain in the European Union or in countries considered to provide equivalent data protection. Revolut says its assistants operate in an isolated environment, with permitted actions defined in advance and a human remaining in control. The presentation addressed data flows but did not discuss potential risks from the model itself or the code it generates.

Because delivery times improved more than coding times, Revolut’s analysis suggests that writing code was not the main bottleneck. Abitbol said engineers also use AI for supporting tasks, including searching email and internal tools. The bank has not abandoned US models: open models are now its priority, but Claude Code, Codex and Gemini remain available for some uses.

Sources: Numerama.