Revolut drops Claude in favor of a Chinese self-hosted model, cuts coding costs by 90%
The switch comes with unlimited tokens.
Image by Bangla press via Shuttertock
Follow us
- Revolut moved about 2,000 engineers from Claude Code and Codex to the self-hosted GLM 5.3 model.
- The company says coding AI costs fell by about 90% after the switch.
- GLM 5.3 coincided with 60% better coding time and 70% better lead time.
- Revolut says it runs public model weights on US infrastructure with human review.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
UK-based neobank Revolut has moved approximately 2,000 engineers to a self-hosted, open-weight model developed in China. This changed how the engineers behave.
The self-hosted GLM 5.3 model, developed by Beijing-based company Z.ai, replaces Anthropic’s Claude Code and OpenAI’s Codex as a default model, according to David Abitbol, who runs AI operations and infrastructure at Revolut.
During his appearance at the AI Engineer conference in New York, Abitbol said the company now spends about a tenth of what it used to pay for American AI labs.
He described the self-hosted GLM model as a way to give engineers more freedom to use AI without unpredictable bills, since they previously had to save limited, expensive tokens for difficult coding tasks.
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
According to Abitbol, the GLM 5.3 uses 56 Nvidia B200 GPUs to serve roughly 2,000 engineers, with a blended cost of nearly $0.30 per million tokens.
Self-hosted models run on users’ own infrastructure, such as physical servers. They offer more predictable pricing because users don’t need to pay third-party providers a margin on every API request.
Cheaper tokens change behavior
Revolut says the benefits of model switching go beyond lower coding bills.
The GLM 5.3 coincided with roughly 60% better coding time and 70% better lead time, which refers to the duration from when a feature request is made until it is delivered to the end user.
This may be because engineering delays stem from work outside writing code, such as inboxes, Jira, and Confluence. Therefore, cheaper access to AI may broaden the scope of tasks people are willing to delegate.
Abitbol admits that there is “a lot of stigma” in using Chinese open-weight models, especially for an EU-regulated bank such as Revolut.
However, he says the company uses only publicly available model weights and runs them on US infrastructure, always with a human in the loop.
Revolut isn’t the first European company eyeing Chinese models
Christian Kroll, the CEO of German search engine Ecosia, said the company was ending the collaboration with France’s AI lab, Mistral, citing quality issues and its environmental commitments. Instead, Ecosia will explore open-weight Chinese models, including GLM.
