On November 3, several outlets reported a talk by David Abitbol, who leads AI operations at Revolut, at the AI Engineer NYC conference on October 13: the fintech company has switched the default coding assistant for about 2,000 engineers to a self-hosted open-weight GLM model from Zhipu, with Claude Code and Codex kept only for exceptions. According to Abitbol, the all-in cost is about $0.30 per million tokens, roughly a tenth of what the company previously paid for OpenAI and Anthropic services.
Reports say the current deployment is GLM 5.3, running on 56 Nvidia B200 GPUs. The system went live in April 2026, and engineers are no longer limited in token usage. Revolut previously gave each person a $300 monthly allowance for Claude Code, and Abitbol said efficiency gains under that cap were limited. After the switch, internal samples show coding time improving by about 60% and lead time from task assignment to delivery improving by about 70%.
The numbers look good, but cause and effect need separating. The drop in time to first token from 14.26 seconds to 1.68 seconds, and a 76% rise in requests per second on the same GPUs, came from tuning inference settings, not from swapping the model alone. Part of the lead-time gain probably reflects engineers no longer rationing tokens and handing surrounding work, such as searching documents and handling tickets, to the assistant. All figures are Revolut's internal statistics, unverified by third parties, and the $0.30 all-in cost does not include hardware purchases or the people running the system.
The part that cannot move
Self-hosting does not replace everything. GLM handles only text and code; image tasks still go to hosted services such as Gemini, and complex tasks return to Claude Code and Codex. As a regulated bank, Revolut downloads the weights onto infrastructure it controls, so inference traffic does not pass through the model provider, and it constrains agents inside a human-in-the-loop sandbox. The safety performance of open models needs separate assessment too: GLM-5.3's attack performance in a cybersecurity evaluation has already been documented independently.
For other companies, the lesson is not that GLM is cheaper than frontier models, but that self-hosting only pays when usage is large enough and tasks are concentrated enough. Without a team tuning 56 GPUs and sandbox governance, money saved on the token bill can easily be given back on the operations bill.
