Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted

Revolut says it cut AI coding inference costs by about 90 percent after shifting its default coding model from commercial services such as Claude to self-hosted open-weight models. The fintech first restricted engineers’ token use because some were spending $1,000 to $10,000 monthly with unclear returns; at roughly $0.30 per million tokens, self-hosted GLM cost about one-tenth as much as OpenAI Codex or Claude Code. After rolling AI tools out to around 2,000 engineers, Revolut reported coding time improved by 60 percent and lead time by 70 percent. Cheap, plentiful tokens also let engineers use AI for debugging and handling Jira and Confluence work, which the company found more valuable than code generation alone.

Revolut now serves GLM-5.3 on 56 Nvidia B200 GPUs, after optimizing quantization, model settings, context lengths, and the separation of prompt processing from generation. These changes matter more than simply adding hardware: the company targets a roughly 256,000-token context, sub-three-second response start, and about 200 generated tokens per second, rather than maximizing context or throughput. GLM is text-only, so a model router sends visual tasks to services such as Gemini or GPT while keeping coding on the self-hosted model. Revolut says this approach makes costs predictable and favors on-premises or regional deployment; it also recognizes privacy and security concerns around Chinese models, especially when accessed through APIs hosted in China.