Revolut cuts AI coding costs tenfold by dropping Claude for GLM models

British fintech company Revolut has drastically optimized its artificial intelligence expenses for software development. David Abitbol, head of operations and infrastructure at Revolut, shared these insights during the AI Engineer conference in New York.

Over the past year, the company managed to cut AI request processing costs by nearly 90%. The primary driver of this massive savings was shifting away from third-party commercial services and deploying proprietary GLM models directly on its own infrastructure.

• Revolut reduced AI coding costs by 10 times by adopting proprietary GLM models.
• AI integration for roughly 2,000 engineers reduced coding time by 60%.
• Overall task completion deadlines dropped by 70% due to process automation.
• The setup runs on 56 Nvidia B200 accelerators with advanced NVFP4 quantization.

Transitioning to GLM models and moving away from Claude

Initially, the fintech enterprise experimented with various cloud-based tools, including Gemini and Anthropic Claude Code. However, strict token consumption limits in commercial products hindered engineers from integrating AI into daily workflows.

The situation shifted dramatically after deploying GLM-5.1 and subsequently the upgraded GLM-5.3 version. The cost of processing one million input and output tokens dropped to about $0.30, roughly ten times cheaper than mainstream commercial alternatives.

Impact on developer productivity and workflows

Substantial cost reduction completely changed how developers approach AI. Previously forced to ration requests, engineers can now freely assign model tasks such as debugging code, breaking down Jira issues, and managing Confluence documentation.

According to company representatives, the bottleneck of development has long been supporting operations rather than writing raw code. Automating these secondary tasks generated the lion's share of the overall time savings for roughly 2,000 engineers.

Technical infrastructure and optimization details

To efficiently run GLM-5.3, the company utilized a cluster of 56 modern Nvidia B200 accelerators. Beyond raw compute power, engineers tuned model parameters, applied NVFP4 quantization, and separated context processing from generation.

The working context was intentionally capped at roughly 256,000 tokens, proving that million-token windows are largely redundant for daily developer tasks. Following optimization, the system delivers the first token in under three seconds at ~200 tokens per second.

Hybrid approach and security considerations

GLM-5.3 did not replace all other models entirely. Since it operates exclusively on text, it cannot handle tasks requiring user interface image analysis, prompting the system's router to direct such queries to Gemini or GPT instead.

Regarding security concerns around Chinese models, leadership views API calls routed through China as the primary risk. Running models locally or within regional cloud resources significantly minimizes these enterprise-grade apprehensions.

What this means for the tech industry and business

Revolut's experience serves as a compelling case study of how major financial institutions can break dependency on expensive cloud APIs. Building optimized, self-hosted AI infrastructure unlocks immense cost efficiency for modern engineering teams worldwide.