StudioRevolut cuts AI coding inference costs about 90% by self-hosting GLM
Original title: British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted
Revolut says it cut AI coding inference costs by roughly 90% in under a year by making self-hosted open-weight models the default instead of relying mainly on commercial services like Claude. After deploying Gemini in early 2025 and Claude Code in early 2026, the fintech moved to Z.ai's GLM-5.1 and then GLM-5.3, served on 56 Nvidia B200 GPUs for about 2,000 engineers, at roughly $0.30 per million tokens, about a tenth the cost of OpenAI Codex or Claude Code. The biggest gains came from debugging and handling Jira and Confluence tickets rather than writing code; with NVFP4 quantization and disaggregated prefill and decode, it achieves sub-3-second time to first token and about 200 tokens per second on ~256k contexts. Since GLM is text-only, a model router sends vision queries to Gemini or …
The case shares concrete cost, performance and optimization details behind replacing commercial coding APIs with self-hosted open models, useful for readers weighing their own deployment options.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.