Revolut Dumps Claude for Chinese AI Model and Slashes Coding Costs by 90%

Revolut has made a major move in the artificial intelligence race, ditching Anthropic’s Claude Code and OpenAI’s Codex as its default coding tools for a Chinese-developed AI model. The result, according to the fintech company, is a staggering 90% reduction in AI coding costs.

The UK-based digital banking giant has reportedly shifted approximately 2,000 engineers to GLM 5.3, an open-weight model developed by Beijing-based AI company Z.ai. Instead of paying external providers for every request, Revolut now operates the model on its own infrastructure, giving developers access to AI assistance without the same unpredictable usage bills.

The move could offer a glimpse into how major companies are rethinking the economics of artificial intelligence as demand for AI-powered software development continues to rise.

Why Revolut Walked Away From Claude and Codex

Revolut’s decision was driven largely by the growing expense of commercial AI services.

David Abitbol, who leads AI operations and infrastructure at the company, discussed the transition at the AI Engineer conference in New York. He explained that Revolut had initially placed strict limits on AI usage because of concerns about the cost of running large-scale coding workloads.

Some developers were spending thousands of dollars each month on AI tokens, creating uncertainty around the financial benefits of deploying these tools across engineering teams.

The switch to GLM 5.3 changed that equation. Revolut says its AI inference costs have fallen to about one-tenth of their previous level, letting engineers use AI more often without constantly worrying about usage limits.

The reported blended cost is around $0.30 per million tokens.

2,000 Engineers and 56 Nvidia GPUs Power the Operation

Revolut’s self-hosted AI infrastructure reportedly runs on 56 Nvidia B200 graphics processing units.

These high-performance chips support the company’s deployment of the GLM model for approximately 2,000 engineers. By hosting the model itself, Revolut can manage computing capacity and usage rather than relying entirely on external AI providers.

Self-hosting also offers more predictable spending. Instead of paying commercial API rates for every interaction, companies invest in their own computing resources and manage the associated operating costs.

However, the savings do not mean running AI is free. Hardware acquisition, electricity, infrastructure maintenance and specialist engineering resources all contribute to the overall expense.

The 90% reduction should therefore be understood as a reported decrease in AI inference costs, not necessarily a 90% cut in Revolut’s entire AI budget.

AI Productivity Reportedly Improves by 60% and 70%

The financial savings are only part of Revolut’s claim.

According to Abitbol, the transition coincided with a 60% improvement in coding-time productivity and a 70% improvement in lead-time productivity.

Lead time measures how long it takes for a feature or requested change to move from the initial request to delivery.

The gains suggest that cheaper AI access can influence more than the speed at which developers write code. Engineers may also be more willing to delegate routine work, documentation, and other tasks that previously received less attention because of usage costs.

With fewer restrictions on AI consumption, developers can potentially incorporate assistance into a wider range of daily activities.

Why a European Bank Is Betting on Chinese AI

The decision is particularly notable because Revolut is a major European fintech operating in a heavily regulated financial environment.

Abitbol acknowledged that using Chinese-developed AI models can attract scrutiny, particularly around security and data privacy.

However, Revolut says it uses publicly available model weights and operates the system on US-based infrastructure, with human oversight remaining part of the process.

The company is not simply sending its engineering workload to an external Chinese AI service. Its self-hosted approach gives it greater control over deployment and infrastructure.

The move also reflects growing interest among European businesses in open-weight AI models, including those developed by Chinese companies.

Could Self-Hosted AI Challenge Claude and OpenAI?

Revolut’s experiment highlights an important shift in enterprise AI adoption. Companies are increasingly weighing the cost of premium commercial models against the economics of running open-weight alternatives themselves.

Frontier models may still be valuable for complex reasoning and demanding tasks, but businesses do not necessarily need their most expensive AI option for every coding request.

If Revolut’s reported savings and productivity improvements hold up at scale, other companies could reconsider how much they spend on commercial AI subscriptions and APIs.

The bigger question is no longer simply which AI model performs best. It is whether businesses can achieve comparable results at a fraction of the cost.

For Anthropic and OpenAI, Revolut’s decision offers another reminder that enterprise customers are increasingly focused on the economics of AI, not just its capabilities.