Revolut Slashes AI Coding Costs via Self-Hosting Pivot
AI-generated image · US National Wire
The fintech firm cut inference costs by 90% by moving away from commercial services like Claude to open-weights models.
Revolut has significantly reduced its AI coding inference costs by transitioning from commercial services to self-hosted models, according to reporting from The Register. David Abitbol, ops and infrastructure lead at the fintech company, stated at the AI Engineer conference in New York that the company slashed these costs by roughly 90% in less than a year.
Prior to the shift, Revolut experimented with Claude Code in early 2026 and Gemini in early 2025. However, Abitbol noted the company was skeptical of the ROI, citing instances where engineers spent between $1,000 and $10,000 per month. To curb these expenses, Revolut initially implemented strict budget limits on tokens.
To scale its AI capabilities for roughly 2,000 engineers, Revolut adopted GLM-5.1 and later GLM-5.3, developed by Z.ai. The company serves these models using 56 Nvidia B200 GPUs. At $0.30 per million input/output tokens, self-hosting GLM 5.1 cost approximately one-tenth of using Anthropic's Claude Code or OpenAI's Codex.
This shift to cheaper tokens allowed engineers to utilize AI for more time-consuming tasks, such as debugging and managing Jira and Confluence tickets, rather than saving tokens for high-difficulty tasks. Abitbol reported a 60% improvement in coding time and a 70% improvement in lead time.
While GLM is the primary driver for code generation, Revolut continues to use a model router to direct vision-based UI queries to GPT Luna or Gemini. Abitbol emphasized that open source is now a priority for the firm, noting that the move has eliminated "bill shock" at the end of the month.