Revolut says it cut its AI coding bill by 90%. What we would copy, and what we would not
Revolut moved about 2,000 engineers off Claude Code and Codex and onto an open model it hosts itself. It says the bill fell to a tenth. The headline is the GPUs. The useful part is what happened when it stopped rationing.
What Revolut reported
David Abitbol, who leads AI operations at Revolut, laid it out at AI Engineer NYC in October. Everything below is Revolut's own account.
- The setup. GLM, an open-weight model from China's Z.ai, on about 56 Nvidia B200 GPUs. Self-hosting started in April 2026. Coverage disagrees on the exact GLM version, so we are not pinning one.
- The cost. About 30 cents per million tokens, blended, and an invoice roughly a tenth of what it had been paying Anthropic and OpenAI.
- The speed. Coding about 60% faster and lead time (picking up a task to merging it) about 70% faster, on a sample of 2,000 engineers. Gemini alone in early 2025 gave about 20% and 10%. Early Claude Code under strict per-engineer budgets showed little impact.
- Where the extra came from. Abitbol put the bigger lead-time gain on non-coding work, like inbox triage and meetings, that engineers handed to AI once tokens stopped being rationed.
- The guardrails. Every agent run is sandboxed with allow-lists, and a human approves. Inference stays in countries covered by a data equivalency framework (the US, EU, UK and Israel were named), with providers chosen for zero data retention by default.
- What stayed on frontier models. Image tasks still go to Gemini or similar, because the self-hosted path is text-only. Claude Code and Codex are now the exceptions, not the default.
What does not travel
The 90% compares token cost with earlier invoices. Reporting on the talk says it leaves out cloud spend, staffing and monitoring, and Revolut has not published total cost. The speed numbers are Revolut's own comparison, and how they were measured was not explained.
The model switch and the end of token limits also happened together, so nobody can say which one did the work. And out of the box, performance was poor. It took heavy tuning (the team reported an 88% gain) and time to first token fell from over 10 seconds to 1.6.
What a small team can copy
- Stop rationing the routine work. Under tight budgets, Revolut's engineers saved the AI for hard tasks and the gains stayed small. A cheap model used on everything beat a strong one used sparingly.
- Route by task. A cheap or open model for routine work, a frontier model for the hard problems. Revolut does exactly this.
- Ask your AI vendor where inference runs and what they keep, and get it in writing. Revolut made both part of the setup, not an afterthought.
- Keep a sandbox and a human approval in the loop. That is what makes it defensible for a regulated bank, and it is good practice for anyone.
Our take, from Nairobi
AI pricing is in dollars and most Kenyan businesses earn in shillings, so a metered plan hurts more here than it does in London. But 56 B200s for 2,000 engineers is not an SME play. For most teams the realistic move is renting open models from a hosted provider, and only self-hosting once volume or compliance justifies it. Do the full sums, GPUs, people and monitoring included, before you copy the headline number.