Revolut says self-hosted GLM cut its AI coding bill by 90 percent

Revolut pays up to 10,000 US dollars per engineer per month for AI coding tools, David Abitbol said last month. Abitbol, who leads AI ops at Revolut, was speaking at the AI Engineer conference in New York on 13 October. The Register's Tobias Mann wrote it up on 3 November.

Revolut has about 2,000 engineers using AI coding tools, Abitbol said. The company started with Gemini in early 2025 and Claude Code in early 2026. The per-engineer bill was between 1,000 and 10,000 US dollars a month at that scale. In the spring of 2026, Revolut switched to a self-hosted open model. Abitbol reports the blended cost is about 0.30 US dollars per million tokens. He said that is a tenth of the commercial price.

The setup

Revolut rents 56 Nvidia B200 GPUs from Together AI. That is seven DGX B200 systems. The model Abitbol described in the talk is GLM 5.3, with 754 billion total parameters and 40 billion activated. Revolut used GLM 5.1 first, then moved to 5.3. The blended cost figure is Revolut's own, reported through The Register.

Why a bank picked an open model

Banks store data they cannot send to a cloud service in the United States. The regulated-data constraint is one reason. The Register notes the cost reduction is another. Abitbol is quoted on a third. "Long context is a scam. You really need to be mindful of that," he said. Revolut's deployment uses a 256,000-token context window, although GLM 5.3 supports a million. Abitbol said tuning cut time to first token by 88 percent on the same hardware. He said it also raised requests per second by 76 percent.

What changed for the engineers

Abitbol told the conference what the move did for the engineers. On the same sample of work, Revolut saw coding time down 60 percent. Lead time was down 70 percent. The Register reports these as Revolut's own numbers, measured by Revolut, from its own engineering telemetry. The method was not shared in detail. But the before-and-after sits inside one company with 2,000 engineers, which is a bigger sample than most published studies.

What a reader building in this space reads from it

The commercial model vendors sell per-token. A per-token bill rations an assistant. An engineer who has 1,000 dollars of model budget this month asks the model fewer questions. An engineer on an unlimited self-hosted service asks it every time. That is where Revolut's coding-time and lead-time numbers come from, per Abitbol's talk. The gain does not come from GLM being faster or better than Claude or Codex. It comes from the budget pressure on the engineer being gone.

A finance team that has data residency rules and an infrastructure team will read this as a template. A ten-engineer startup will not: 56 B200 GPUs are not a reasonable starter setup. For most teams, the question is simpler. Is a per-token bill making your engineers skip the assistant?

Source

British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted, Tobias Mann, The Register. Shipping like a startup, regulated like a bank, AI Engineer NYC 2026, David Abitbol.