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

British fintech company Revolut has significantly reduced its AI coding inference costs by 90% by moving to self-hosted models and cutting out commercial services like Claude. Ops and infrastructure lead David Abitbol reported a 60% improvement in coding time and a 70% improvement in lead time after implementing AI coding tools for around 2,000 engineers.

Initially, Revolut used GLM-5.1, which they upgraded to GLM-5.3, serving it with 56 Nvidia B200 GPUs. Revolut's skepticism about AI costs led them to limit engineer access to tokens, driving them to self-host models. With self-hosting, GLM 5.1 costs about a tenth of using OpenAI's Codex or Anthropic's Claude Code. The biggest gains came from allowing AI to handle tasks like debugging and Jira tickets, as unlimited tokens freed engineers from token-saving constraints.

Revolut optimized GLM-5.3 by tuning hyperparameters, implementing NVFP4 quantization, disaggregated prefill and decode, and adjusting context lengths. They found that a 256,000-token context window is sufficient, with sub-3-second time to first token and 200 tokens per second generation speed being optimal. GLM-5.1 is used for code generation, while Gemini or GPT Luna handles vision queries, following a model router.

While open-source models like GLM offer cost advantages, Revolut acknowledges privacy and security concerns, noting that running models on-prem or regional cloud resources mitigates these issues.