Revolut Achieves 90% Reduction in AI Coding Costs with Self-Hosted Models

Revolut, the British fintech company, has announced a remarkable reduction in its AI coding inference costs, achieving a decrease of around 90% in less than a year. This significant cost-saving measure was accomplished by shifting to self-hosted, open-weight models, moving away from reliance on commercial services such as Anthropic’s Claude.

At the recent AI Engineer conference in New York, David Abitbol, Revolut’s operations and infrastructure lead, shared insights into the company’s transition. Following the rollout of AI coding tools to approximately 2,000 engineers, Revolut reported a 60% improvement in coding time and a 70% enhancement in lead time, which measures the duration from project initiation to delivery.

The self-hosted models, sourced from the Chinese AI lab Z.ai, began deployment with the GLM-5.1 model in April 2026 and later upgraded to GLM-5.3. The new model operates on 56 Nvidia B200 GPUs within Revolut’s infrastructure, with a cost of about $0.30 per million tokens. Prior to this switch, the company faced substantial monthly bills ranging from $1,000 to $10,000 due to usage of commercial APIs, prompting strict budget controls and token rationing.

By self-hosting, Revolut eliminated the need for token rationing, allowing engineers to focus on more complex tasks rather than mundane chores like documentation and ticket filing. The company also optimized its system post-launch, achieving a 76% increase in throughput and reducing the time-to-first-token response to an average of 1.6 to 1.68 seconds.

Revolut’s decision to self-host AI models is particularly noteworthy given its status as a regulated financial institution. This approach ensures compliance by running AI weights on US infrastructure while maintaining human oversight of code suggestions. The self-hosting model requires significant upfront investment in hardware and expertise, but with a workforce of 2,000 engineers, the economics become favorable.