Revolut has moved around 2,000 engineers from Claude Code and OpenAI’s Codex to a self-hosted model developed by Chinese AI company Z.ai.
The switch has brought the company’s AI coding costs down to roughly a tenth of their previous level.
David Abitbol, who oversees AI operations and infrastructure at British Fintech, discussed the move at the AI Engineer conference in New York in October. Revolut now uses Z.ai’s GLM 5.3 as its default coding model. It runs on 56 Nvidia B200 GPUs and costs the company around $0.30 per million tokens.
Before the switch, some engineers were spending between $1,000 and $10,000 a month on commercial AI tools. Revolut responded by setting strict limits on token usage. Developers had to be selective about when they used the more expensive models, often saving them for complex coding tasks rather than routine work.
The self-hosted model changed that. With costs no longer tied to individual usage in the same way, Revolut removed the limits. Engineers began using AI for everyday tasks too, including sorting through Jira tickets, searching Confluence and handling email.
Abitbol said the change helped improve coding time by around 60%. Lead time, the period between an engineer picking up a task and getting it ready for deployment, fell by roughly 70%. He attributed much of the improvement to engineers being able to use AI more typically, rather than to a major difference in the model’s coding abilities.
Using a Chinese-developed model in a regulated European financial business has raised some questions. Abitbol acknowledged that Chinese AI models still carry a stigma in this environment. Revolut’s approach is to download the publicly available model weights and run the model on U.S. infrastructure, so data is not sent to servers in China.
The company also keeps engineers involved in the process. AI tools assist with development, but people remain responsible for reviewing and managing the work.
Claude Code and Codex are still available when engineers need them. However, GLM 5.3 is now the default for day-to-day work. Revolut began self-hosting earlier versions of the model in April before moving to the current release.
