London-based fintech Revolut says it has cut the cost of its AI coding tools by about 90 percent in under a year. The trick: it made a self-hosted, open-weights model its default instead of paying per-token fees to services such as Anthropic’s Claude Code or OpenAI’s Codex.
The details come from a talk by David Abitbol, Revolut’s ops and infrastructure lead, at the AI Engineer conference in New York in October, as reported by The Register on Nov. 3.
What Revolut Did
Revolut tried Google’s Gemini in early 2025 and Claude Code in early 2026. According to The Register, some engineers were spending $1,000 to $10,000 a month with little to show for it, so the company capped budgets. The caps left engineers rationing tokens, which pushed the team to host a model itself.
It started with GLM-5.1 and moved to GLM-5.3, both from China’s Z.ai. GLM-5.3 runs on 56 Nvidia B200 GPUs, the equivalent of seven eight-GPU DGX B200 servers, serving roughly 2,000 engineers. Revolut puts the cost at about $0.30 per million tokens, around a tenth of what it paid for commercial tools.
Abitbol said most of the gains came from tuning rather than buying more hardware: adjusting hyperparameters, using NVFP4 quantization (a lower-precision number format), and splitting the prefill and decode stages of inference.
Why It Matters
With budget limits gone, Revolut says engineers now use AI for debugging and for handling Jira and Confluence tickets, not just writing code. The company claims a 60 percent improvement in coding time and a 70 percent improvement in lead time.
For US businesses, the takeaway is less “switch models” and more “check your meter.” Per-token pricing makes heavy AI use unpredictable, and a company with enough steady volume and engineers to run the hardware may find owning the infrastructure cheaper.
The Catch
The cost and productivity figures come from Revolut’s own conference presentation, and The Register did not verify them independently. The savings also leave out the engineering staff needed to run the cluster.
GLM is text-only and not ideal for interface design work, so Revolut routes image-related requests to other models such as Gemini or GPT. Abitbol also acknowledged security concerns about Chinese open-weights models. He said the main risk is reaching them through APIs hosted outside China, and that running them on-premises or in a regional cloud greatly reduces it.
The story also lands amid worries about AI-generated code quality. Surveys have linked AI coding tools to outages at most big software teams, a reminder that cheaper tokens do not remove the need for review.
More on Contoh
- AI Coding Tools Are Linked to Outages at Most Big Software Teams, Surveys Find
- Claude Opus 5.5 Is Cutting Off Reverse-Engineering Work. Here’s the Likely Reason