UK-based digital bank Revolut was able to cut the inference costs for its AI coding by around 90 per cent through switching from OpenAI and Anthropic to a self-hosted, open source model, the company said.

The news was first reported by The Register and initially revealed in a speech given by David Abitbol, operations and infrastructure lead at the firm, at a conference in New York.

Revolut began its AI rollout by deploying Gemini and Claude Code, but found it lacked a clear return on investment, he said: “We were highly sceptical about the cost of AI. We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI, and so when we launched it we launched it with strict budget limits.”

This meant that engineers had access to a limited amount of tokens, which left engineers saving them for the most complex tasks rather than incorporating AI use into their general workflows, Abitbol continued.

This led it to test self-hosted GLM models developed by Chinese company Z.ai. It initially adopted GLM-5.1 before moving to GLM-5.3, he added.

This second model coincided with a roughly 60 per cent decrease in coding time and a 70 per cent decrease in lead time, as well as significantly lower costs per token.

Deploying GLM-5.3 across the organisation required 56 of Nvidia’s advanced B200 GPUs, Abitbol said. The largest gains did not come from increasing the number of GPUs, however, but from tuning the model’s hyperparameters and other software optimisations.

The open-source, self-hosted model is working well for Revolut, Abitbol said. Although he acknowledged security concerns around Chinese models, he argued that these are largely important when accessing them through APIs running in China. Self-hosting solves this problem, he said.

This is the latest example of a large financial institution choosing to alter its AI strategy in response to cost concerns. In October, Santander announced that it had moved from large, general-purpose AI models towards smaller, specialised ones, reducing costs and increasing output quality.

On Wednesday, it was revealed that Goldman Sachs has restricted most traders and bankers from using advanced AI coding tools after a review raised concerns around productivity and risk management.

---