Revolut says it has cut its AI coding bill by roughly 90 percent in less than a year, by making self-hosted open-weight models the default for its roughly 2,000 software engineers instead of relying mainly on commercial coding services. The British fintech described the switch at the AI Engineer conference in New York last month, where its operations and infrastructure lead, David Abitbol, said the bill is now about a tenth of what the company used to pay Anthropic and OpenAI. Frontier models remain available, but only as the exception.
The switch, in stages
Revolut did not jump straight to self-hosting. It first deployed Google’s Gemini in early 2025, then Anthropic’s Claude Code in early 2026, launching the latter with strict per-engineer budget limits. Abitbol said the company was highly sceptical about AI costs, having seen individual engineers spend $1,000 a month, and in some cases $10,000 a month, without a convincing return.
The turning point came when Revolut began hosting models itself. It first adopted GLM-5.1, an open-weight model from the Chinese lab Z.ai, and has since moved to GLM-5.3, which it now serves on 56 Nvidia B200 graphics processors. Because the weights are published, Revolut runs them on its own infrastructure in the United States rather than through Z.ai’s hosted service.
At about $0.30 per million input and output tokens, self-hosting GLM cost roughly a tenth of using OpenAI’s Codex or Anthropic’s Claude Code. The model is large: roughly 754 billion total parameters, with about 40 billion activated for each token generated. It needs around 380 gigabytes of graphics memory just to hold the weights, plus more per user for context.
What changed for engineers
The headline productivity figures are striking. Abitbol reported roughly a 60 percent improvement in coding time and a 70 percent improvement in lead time, measured from picking up a task to merging it. Those are Revolut’s own figures, and the comparison does not isolate the model change from other causes, such as the removal of token limits.
The bigger gain, Abitbol argued, was not in writing code at all. With tokens effectively unlimited, engineers stopped saving them for the end of the quarter or the hardest task, and spent them instead on debugging and routine tickets. The bottleneck, he said, was never really code generation; it was the surrounding work engineers previously did by hand.
Tuning mattered as much as hardware. Abitbol said the greatest performance gains came not from adding graphics processors but from tuning hyperparameters, switching to NVFP4 quantisation, and separating prefill and decode work. The result was a sub-three-second time to first token on context windows of around 256,000 tokens, with generation speeds of about 200 tokens a second.
The limits of the claim
GLM has not replaced the frontier models entirely. The model is text-only, so it is a poor fit for user-interface work, and Revolut routes vision queries to other models such as Gemini or GPT Luna. Claude Code and Codex also remain available for exceptional cases, so the switch is best read as a change of default, not a clean break.
The 90 percent figure needs the same care. It describes AI coding inference costs, and analyses of the talk note that it excludes cloud, staff and monitoring costs, which sit outside the token price. The $0.30 per million tokens is a blended deployment cost at Revolut’s utilisation, not a price another company can simply copy.
Security questions follow the model choice. Abitbol acknowledged that Chinese open-weight models can raise privacy and security concerns, but argued the real risk lies in using interfaces hosted in China. Running published weights on infrastructure in the United States, he said, leaves nothing to worry about, a claim regulators and customers will test.
Our read
Revolut’s story is less about one model than about a pricing model breaking. When coding assistants billed per token through an interface, every engineer became a cost centre to be capped, and the caps quietly throttled the productivity the tools promised. Self-hosting flips the incentive: the marginal token is nearly free, so engineers spend tokens freely on unglamorous work.
That makes Revolut a useful test case, but a hard one to copy. Its economics depend on scale, with 2,000 engineers to amortise 56 top-end processors, and on engineering depth to tune a 754-billion-parameter model into a responsive service at bank scale.
Smaller firms will more likely rent the same idea through inference providers than replicate it. That is why the broader coding-tools market still matters more than any single deployment, from Cursor’s point-and-fix design mode to Claude Code’s recent guardrail changes.
The strategic signal is aimed squarely at Anthropic and OpenAI. If a major customer can serve a frontier-class open model at a tenth of the price and keep the frontier only for exceptions, the subscription and usage-based model for coding tools faces a ceiling. The labs’ answer so far has been capability: keep the hardest tasks exclusive to the frontier.
Why It Matters
This matters because coding was the first place AI agents proved their economic value, and Revolut is the first large, regulated company to show its working in public. If its numbers hold, technology chiefs everywhere gain a credible alternative to per-token billing, and a reason to ask vendors why their prices should stay where they are.
It also sharpens the open-weight question for every bank and regulated firm watching. Revolut’s answer, that published weights run on home infrastructure are safer than foreign-hosted interfaces, is a template others may follow. The next year will show whether regulators accept that distinction, or whether the origin of the weights matters regardless of where they run.
Background
Revolut is a British financial technology company whose engineers previously moved from Gemini in early 2025 to Claude Code in early 2026 before adopting GLM, the open-weight model family from Z.ai, also known as Zhipu. Abitbol laid out the switch in a talk titled “Shipping like a startup, regulated like a bank” at the AI Engineer conference in New York.
Source: Register report and Dealroom report
0 Comments