Revolut swaps Claude Code for a Chinese open-weight model, cutting its AI coding bill by 90%
What's the deal? Revolut has moved its roughly 2,000 software engineers off AnthropicDealroom has a profile for this one. Try Dealroom →'s Claude Code and OpenAI's Codex as the default coding tools. Its engineers now run on GLM, an open-weight model from Chinese AI lab Z.ai, which Revolut hosts itself. According to Revolut, its bill is now about a tenth of what it used to pay Anthropic and OpenAI.
Who said it? David Abitbol, who runs AI operations and infrastructure at Revolut, laid out the numbers in his talk "Shipping like a startup, regulated like a bank" at the AI Engineer conference in New York. The details went viral in early November after Trending Topics wrote them up.
The numbers. Revolut serves GLM to its engineers from about 56 Nvidia B200 GPUs, at a blended cost of around $0.30 per million tokens. In a sample of about 2,000 engineers, Abitbol reported roughly 60% shorter coding time and 70% shorter lead time (from picking up a task to merging it) after the switch. Those are Revolut's own figures, and the company's comparison does not isolate the model change from other causes, such as the removal of token limits.
Why switch? Revolut first rolled out Google's Gemini, then Claude Code with strict per-engineer budgets. Abitbol said the early Claude Code rollout was costly for a modest gain, and that rationed tokens made engineers save AI for hard coding tasks. With a flat, self-hosted bill, the limits are gone. "The bottleneck for engineering is not necessarily coding," he said, pointing to time lost in inboxes, Jira and Confluence.
Chinese weights, Western servers. Abitbol acknowledged the "stigma" of using a Chinese open-source model at a regulated bank. Revolut only uses the open weights and runs them on its own infrastructure in the US, not through Z.ai's hosted service. Image tasks are routed to Gemini, and Claude Code and Codex remain available as the exception for the hardest work. Coding agents run in a sandbox with approval rules, with a human in the loop.
What could go wrong? The 90% saving excludes cloud, staff and monitoring costs, and depends on heavy tuning: Abitbol said out-of-the-box inference was "terrible" until the team cut time to first token from about 14 seconds to 1.7 seconds on the same GPUs. Smaller companies without Revolut's scale and infrastructure team may not get the same result.
The signal: One of Europe's most valuable tech companies is now running a Chinese open-weight model as the default for daily coding work across its whole engineering team, not as a pilot. It is a warning for the AI labs' enterprise business: for routine, high-volume work, "good enough and cheap" is starting to win, with frontier models kept for the exceptions.
Read more: AI Engineer (talk summary and video), Trending Topics
Image credit: Revolut
