Revolut transitioned approximately 2,000 of its engineers to a self-hosted GLM model for coding, moving previously used tools like Claude Code and Codex to an exception list. The system runs on 56 Nvidia B200 GPUs and costs a blended 30 cents per million tokens, which is about 10% of what the fintech firm paid for APIs from OpenAI and Anthropic.
David Abitbol, who leads AI operations at Revolut, stated the transition helped the company stop rationing tokens, allowing engineers to use AI for all tasks instead of only the most difficult tickets. The shift to an internal model reduced engineering lead time by 70%.
