- Revolut moved its coding assistant for about 2,000 engineers from paid Anthropic and OpenAI services to a self-hosted GLM model; the bill fell to about one tenth.
- It runs on 56 B200 GPUs at about $0.30 per million tokens, and after tuning the first reply arrives in 1.68 s instead of 14.26 s.
In short: Revolut stopped paying mostly for outside coding assistants and set up its own server running a model with open weights. According to the company, its AI coding bill dropped to roughly a tenth. David Abitbol, Revolut’s ops and infrastructure lead, shared the details at the AI Engineer conference in New York on 13 Oct 2026; The Register wrote it up on 3 Nov 2026.
What changed
- The assistant serves about 2,000 engineers. Revolut used Gemini earlier and Claude Code from early 2026; according to The Register, some engineers were spending $1,000–10,000 a month.
- The default model is now GLM (the talk mentions versions 5.1 and 5.3) on 56 B200 cards, which are NVIDIA GPU chips.
- The blended cost is about $0.30 per million tokens, and the invoice is about one tenth of the earlier Anthropic and OpenAI spend.
- After tuning on the same cards, requests handled per second rose 76% and time to first token fell 88%: from 14.26 s to 1.68 s.
- The target is under 3 s to the first word and about 200 tokens per second; above 5 s, people start complaining.
- GLM reads text only, so a model router sends questions with screenshots to Gemini or GPT.
How to read this
These numbers come from one company’s talk, not an independent test. The conference page notes that the roughly 60% better coding time and 70% better lead time are an internal comparison that does not separate the model change from other factors. The $0.30 rate depends on large, steady load from thousands of people; a small team with one server will not get that price. My takeaway: your own model pays off at scale and with people who know how to tune it. I write more about downloadable models in the Open models hub and about AI bills in the Model pricing hub.