Revolut makes self-hosted GLM its default AI coding model in place of tools like Claude Code, saying inference costs fell about 90%
The Register reported on 3 November that David Abitbol, ops and infrastructure lead at British fintech Revolut, told the AI Engineer conference in New York last month that the company had, in under a year, made self-hosted Z.ai GLM models its default for AI coding instead of relying mainly on commercial services such as Claude Code, cutting AI coding inference costs by roughly 90%.
Abitbol said Revolut deployed Gemini in early 2025 and Claude Code in early 2026, saw engineers spending $1,000 to $10,000 a month without much return, and launched with strict budget limits that left engineers with a limited number of tokens, which led it to look at self-hosting. By his account, self-hosting GLM 5.1 cost about $0.30 per million input/output tokens, roughly a tenth of using OpenAI’s Codex or Anthropic’s Claude Code.
Revolut started on GLM-5.1 and moved to GLM-5.3, which it now serves on 56 Nvidia B200 GPUs for roughly 2,000 engineers; it reports a 60% improvement in coding time and a 70% improvement in lead time after rolling out AI coding tools. Abitbol said the biggest gains came not from writing code but from letting the models handle debugging and Jira and Confluence tickets once tokens were no longer rationed.
He said the main performance gains came from tuning hyperparameters, switching to NVFP4 quantization, disaggregating prefill and decode and tuning context lengths rather than adding GPUs, and that while a one-million-token window can be served, 256K is what is really needed. With these changes Revolut reports sub-3-second time to first token at around 256K context and about 200 tokens per second.
GLM has not replaced commercial models entirely: it is text-only, so Revolut uses a model router to send vision queries to Gemini or GPT Luna while code generation stays on GLM. On privacy and security concerns about Chinese open-weight models, Abitbol said the main risk is using APIs run out of China, not running the models on-prem or in regional clouds. These are Revolut’s own figures; the report gives no independent accounting.