The time spent writing code fell by 60%, while task completion times dropped by 70%.
British fintech company Revolut has reduced the cost of processing AI coding requests by 90% in approximately one year. The company previously relied mainly on commercial services such as Anthropic’s Claude Code and OpenAI Codex, but has since made its own GLM models, running on its infrastructure, the primary solution. David Abitbol, Revolut’s Head of Operations and Infrastructure, discussed the shift at the AI Engineer conference in New York.
According to Abitbol, after AI tools were introduced for approximately 2,000 engineers, the time spent writing code fell by 60%, while task completion times decreased by 70%. The company initially used Gemini and began experimenting with Claude Code in early 2026. However, strict token usage limits made it difficult for engineers to rely on AI for everyday tasks.
Deploying the GLM-5.1 model cost Revolut approximately $0.30 per million input and output tokens, making it around ten times cheaper than commercial alternatives. The company later moved to GLM-5.3.
The lower costs changed not only the budget but also how engineers worked. They no longer needed to ration their requests and began delegating more tasks to AI, including debugging software, analysing Jira tasks and working with documentation in Confluence.
According to Abitbol, automating these activities accounted for a significant part of the time savings. The main bottleneck in software development was not necessarily writing code, but the additional tasks surrounding the development process.
Revolut uses 56 Nvidia B200 accelerators to run GLM-5.3. However, the performance improvements were not achieved through computing power alone. The company optimised the model's parameters, applied NVFP4 quantisation and separated input-context processing from response generation into different stages of its infrastructure.
Revolut also limited the working context to approximately 256,000 tokens. According to Abitbol, the fact that a model supports a one-million-token context window does not mean developers actually need it. Following optimisation, the system delivers its first token in less than three seconds and generates approximately 200 tokens per second.
GLM-5.3 has not replaced every other model used by the company. It works exclusively with text and is therefore unsuitable for tasks that require analysing images of user interfaces. Revolut's routing system directs such requests to Gemini or GPT, while GLM remains the primary model for code generation.
Abitbol also acknowledged that using Chinese models may raise questions about privacy and security. However, he considers the main risk to be accessing these models through APIs operating in China. In his assessment, these concerns are significantly reduced when the models are deployed on the company's own infrastructure or through regional cloud resources.
