“The billing is flat, we don’t have any bill shock at the end of the month, and the frontier is the exception.” That is how Revolut’s David Abitbol described the result of moving the fintech’s AI coding from commercial services such as Claude Code to self-hosted open models. In less than a year, Revolut cut its AI coding inference costs by about 90%, The Register reports, while about 2,000 engineers saw a 60% improvement in coding time. It is one of the clearest real-world examples yet of a large company bringing AI in-house to control the bill.
Go deeper: why token costs are becoming a strategic problem: The Agent Era Has a Token Bill.
What Revolut did
Revolut deployed Gemini in early 2025 and Claude Code in early 2026, with strict budget limits after seeing engineers spend “$1,000 a month, $10,000 a month with not so much ROI,” Abitbol, its ops and infrastructure lead, told the AI Engineer conference in New York. It then switched to GLM, an open-weights model from Chinese developer Z.ai. At $0.30 per million tokens, self-hosting GLM-5.1 cost about a tenth of Codex or Claude Code. Revolut now serves GLM-5.3 on 56 Nvidia B200 GPUs.
Cheap tokens changed how engineers worked. The biggest gains came not from writing code but from letting AI handle debugging and Jira and Confluence tickets. “When you have unlimited tokens, you’re free to do that,” Abitbol said. Most of the performance came from tuning rather than more hardware, and he argues giant context windows are overrated: “Long context is a scam… Even though you can serve one million tokens, 256 [thousand] is really what you need.” His targets are a first response in under 3 seconds and about 200 tokens a second.
Frontier models haven’t disappeared. GLM is text only, so a router sends image tasks such as UI work to models like Gemini or GPT. On concerns about Chinese models, Abitbol said the main risk is using them through APIs hosted in China, not running the weights on your own or regional infrastructure.
Why it matters
Revolut’s case reflects a shift: the frontier model becomes a premium tool for hard cases, and an owned, flat-cost model handles the volume. The same logic applies beyond coding. Businesses that rent every layer of their stack, from AI to commerce, pay a usage tax that grows with success. Owning the core layer converts that into a predictable cost.
That is the reasoning behind VBWD, the self-hosted subscription and commerce platform we build: your billing, customers and data run on your own infrastructure, with no platform cut on each sale (The Zero-Platform-Cut Economics of Owning Your Stack). For teams adding AI to their product, we have also written about how to add AI to your SaaS without leaking customer data.
Written for Red Robot with AI assistance and human editing. VBWD is our own product. Based on reporting by The Register.