Revolut cuts AI coding costs 90% with self-hosted GLM

Revolut cut its AI coding costs by about 90% by self-hosting GLM in place of commercial coding agents, The Register reported on November 3. The figures come from an October talk at AI Engineer NYC by David Abitbol, who leads AI operations at the fintech. His summary: "our invoice is one-tenth of what we were spending on Anthropic and OpenAI." The setup is not for a ten-person company. The lesson behind it is.

What actually happened

• The switch: Revolut made self-hosted GLM its default coding model. It started on GLM-5.1 and moved to GLM-5.3, per The Register. Frontier models such as Claude Code and Codex stay available "as exception."
• The hardware: 56 Nvidia B200 GPUs serve about 2,000 engineers. The endpoints are OpenAI-compatible, so tools like OpenCode and Codex can point at them.
• The tuning: FP4 quantization (Abitbol said it cut the GPU footprint by about 40% with no loss he could measure), prefill/decode disaggregation, speculative decoding, a cache-hit rate above 90%, and a 256K context cap. In his words, "long context is a scam."
• The cost: a blended $0.30 per million tokens on Revolut's own utilization. That is an internal figure, not a price you can buy.
• The trigger: during the earlier Claude Code rollout, some engineers spent $1,000 to $10,000 a month with little return, so Revolut set strict budgets.
• The gap: GLM is text-only, so image work goes to other models.

Revolut also showed internal productivity gains after the switch. Its own comparison does not separate the model change from other changes, so we do not repeat those numbers as proof.

Why it matters for your business

You will not rent 56 B200s. But two facts here apply at any size.

First, coding-agent spend grows without a ceiling. One engineer can burn $10,000 a month on a frontier agent. If you pay for Claude Code, Codex or Cursor seats for your team, set budgets for each person and read the usage report every month.

Second, a single default model is a choice. Revolut kept frontier models for hard problems and moved routine work to a cheaper model. You can do the same with hosted open-weights APIs and no GPUs: point your coding tools at an OpenAI-compatible endpoint for daily work, and escalate to the frontier model when it fails. The model layer is the part you should be able to swap.

Sources: The Register, AI Engineer NYC talk page.