Revolut has made Zhipu's GLM the default AI coding model for approximately 2,000 engineers, cutting inference costs by roughly 90% in less than a year by replacing commercial tools like Claude Code as the primary option1,2.
David Abitbol, Head of AI Operations at Revolut, said the company previously allocated each engineer a monthly budget of $300 for Claude Code but saw limited overall efficiency gains. Claude Code and OpenAI's Codex now serve as secondary tools reserved primarily for complex tasks.
Since April, Revolut has been running GLM 5.3 in-house on approximately 56 NVIDIA B200 GPUs, bringing costs down to an estimated $0.30 per million tokens. That brought AI expenses to one-tenth of what the company was paying OpenAI and Anthropic.
The cost reduction opened the door to broader usage. Revolut gave engineers unlimited token access, and they now use AI for internal documentation searches, Jira ticket management, and tracking requirement specifications in Slack. The Register noted that cheaper tokens prompted engineers to apply AI to tasks like debugging and tackling Jira tickets.
Internal data cited by Revolut shows a 60% improvement in coding speed and a 70% acceleration in the cycle time from task assignment to delivery.
ANALYSIS The scale of the deployment — 2,000 engineers on a self-hosted open-weight model — and the magnitude of the reported cost drop illustrate a concrete trade-off: accepting a model that may be less capable on the hardest tasks in exchange for dramatically cheaper tokens that drive higher aggregate usage. Revolut's decision to keep Claude Code and Codex on hand for complex work rather than eliminating them entirely reflects that calculus.
The move comes as enterprise engineering teams are grappling with the operational costs of AI-assisted development. Shopify's mobile engineering team recently found that running multiple AI agent workflows in parallel maxed out developer hardware, making infrastructure — not model quality — the binding constraint[2].