Self-Hosted GLM Models

Revolut Replaces Claude Code, Cutting AI Coding Costs by 90%

Revolut, Revolut AI coding, Revolut AI coding costs, How Revolut cut AI coding costs by 90 percent, Revolut switches from Claude Code to open-weight models
Facebook
X
LinkedIn
Reddit
WhatsApp
Source: RixAiArt/Shutterstock.com

Fintech company Revolut runs Z.ai’s open-weight models on 56 Nvidia B200 GPUs. The move has reportedly cut its AI coding costs by around 90 percent.

British fintech company Revolut says it has reduced the cost of AI-assisted software development by around 90 percent in less than a year. The company achieved this by making self-hosted open-weight models its default, rather than relying primarily on commercial services such as Anthropic’s Claude Code and OpenAI’s Codex. David Abitbol, Revolut’s Head of Operations and Infrastructure, shared the details at the AI Engineer conference in New York, as reported by The Register.

Ad

Budget Limits Trigger the Shift

Revolut initially introduced Google’s Gemini in early 2025, followed by Claude Code in early 2026. According to Abitbol, the company was skeptical about the cost of AI and imposed strict spending limits after individual developers had run up monthly bills of USD 1,000 or even USD 10,000 without delivering results to match the expense. By introducing fixed token allowances for developers, Revolut began exploring its own models.

Since April 2026, Revolut has been running GLM-5.1, developed by Chinese AI company Z.ai. The company has since upgraded to version GLM-5.3. The model runs on 56 Nvidia B200 GPUs within Revolut’s own infrastructure.

According to Abitbol, costs amount to around USD 0.30 per million input and output tokens combined. That is approximately one-tenth of what Revolut previously paid for Claude Code and Codex. However, the available reports do not clarify whether this figure fully accounts for the cost of acquiring the hardware. Revolut has since lifted the token limits for its approximately 2,000 developers.

Ad

“Open source is now a priority at Revolut, and our economics are favorable.”

David Abitbol, Head of Operations and Infrastructure at Revolut

Biggest Gains Beyond Writing Code

Following the introduction of AI tools, Revolut reports a 60 percent improvement in coding time and a 70 percent improvement in turnaround time. However, Abitbol sees the greatest benefits not in code generation itself, but in debugging and handling tickets in Jira and Confluence. These figures are based on the company’s own data.

GLM has not completely replaced commercial models. It processes text only and is less suited to tasks such as designing user interfaces. Revolut therefore uses a routing system to direct image-based requests to models such as Gemini or GPT. Claude Code and Codex remain available for exceptional cases.

The cost of AI-assisted coding is becoming a growing concern for businesses. Amazon has instructed its developers not to use token consumption as a measure of productivity. According to reports, Accenture has capped its centrally funded Claude budget at USD 29 per person per month. At the same time, a VentureBeat survey suggests that many companies are leaving their own GPUs largely unused because employees continue to favor closed-source models.

(Editorial Team)

Ad

Artikel zu diesem Thema

Weitere Artikel