Revolut reports lower AI coding costs with self-hosted GLM

Revolut began self-hosting open-source models in April 2026 and chose GLM for coding workloads.[2] The model is an open-weight system from Chinese AI lab Z.ai, and Revolut self-hosts it.[1] The company moved roughly 2,000 engineers off Claude Code and Codex as their default coding tools.[1] That shift is not a complete replacement of commercial AI: Revolut still uses Claude Code, Codex and Gemini for specialised work.[2] Image-related tasks are routed to Gemini or similar hosted models, while the self-hosted path handles text and code generation.[5]

Revolut says its invoice is now one-tenth of what it spent on Anthropic and OpenAI, and that self-hosting GLM brings its blended token cost to 30 cents per million.[2] The reported 90% savings do not include cloud, staff or monitoring costs.[4] In a sample of 2,000 engineers, the company reports coding time fell 60% and engineering lead time fell 70%.[2] The comparison does not isolate the effect of changing models from other factors, including removing token limits.[1] Revolut says removing those constraints reduced time spent searching documentation, managing inboxes and coordinating development tasks.[2]

Revolut runs GLM-5.3 for about 2,000 engineers on 56 NVIDIA B200 GPUs.[2] It says requests per second rose from about three to more than seven between April and August, while time to first token fell from more than 10 seconds to about 1.6 seconds as demand rose 76%.[2] Interactive coding workloads run at about 200 tokens per second, with throughput reaching up to 800 after 5 p.m., when GPU demand is lower.[2] Before adopting GLM, Revolut’s rollout of Gemini improved coding time by 20% and lead time by 10%, according to Abitbol.[4] The company also says it values flat billing and predictable monthly costs.[2]