Revolut swaps Claude for GLM open-weight model, cutting AI coding bill

Revolut has moved its roughly 2,000 software engineers off Anthropic's Claude Code and OpenAI's Codex as the default coding tools. Engineers now run on GLM, an open-weight model from Chinese AI lab Z.ai, which Revolut hosts itself. According to Revolut, the bill is now about a tenth of what it used to pay Anthropic and OpenAI. dealroom.co's report records the switch and the figures behind the new default.

This piece is for engineering leaders and developers who run coding agents across a large team. It covers Revolut's path from Google's Gemini and budgeted Claude Code to a flat, self-hosted GLM default. 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.

Revolut swaps Claude for GLM: what actually changed

Revolut first rolled out Google's Gemini, then Anthropic's Claude Code with strict per-engineer budgets. David Abitbol said that early Claude Code rollout was costly for a modest gain, and that rationed tokens made engineers save AI for hard coding tasks. The default is now GLM, the open-weight model from Chinese AI lab Z.ai, hosted by Revolut for roughly 2,000 software engineers. OpenAI's Codex and Claude Code remain the exception for the hardest work, and image tasks are routed to Gemini.

In a sample of about 2,000 engineers, Abitbol reported roughly 60% shorter coding time and 70% shorter lead time after the switch. Lead time runs from picking up a task to merging it. Those are Revolut's own figures, and the comparison does not isolate the model change from other causes, such as the removal of token limits. With a flat, self-hosted bill, the limits are gone. The bill is about a tenth of what Revolut used to pay Anthropic and OpenAI, a 90% saving.

Revolut swaps Claude for GLM: how it works

Revolut serves GLM to its engineers from about 56 Nvidia B200 GPUs, at a blended cost of around $0.30 per million tokens. It uses only the open weights and runs them on its own infrastructure in the US, not through Z.ai's hosted service. Image tasks go to Google's Gemini. Claude Code and Codex stay available for the hardest work. Daily coding for the engineering team goes to the self-hosted GLM service, and that bill is flat, replacing the old per-engineer token budgets.

Coding agents run in a sandbox with approval rules, with a human in the loop. Speed depended on heavy tuning by Revolut's infrastructure team. 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. Those machines are the Nvidia B200s, about 56 of them, running Z.ai's open weights in the US. The 90% saving depends on that tuning. The blended cost is around $0.30 per million tokens.

Revolut swaps Claude for GLM: why it matters

"The bottleneck for engineering is not necessarily coding," Abitbol said, pointing to time lost in inboxes, Jira and Confluence. Strict per-engineer Claude Code budgets led engineers to save AI for hard coding tasks. Abitbol said that rollout was costly for a modest gain. A flat, self-hosted bill removes the limits. For about 2,000 engineers, Revolut's figures show roughly 60% shorter coding time and 70% shorter lead time from task pickup to merge, and the comparison does not isolate the model change from other causes.

Revolut, one of Europe's most valuable tech companies, now runs a Chinese open-weight model as the default for daily coding across its whole engineering team, not as a pilot. For the AI labs' enterprise business, that is a warning. For routine, high-volume work, "good enough and cheap" is starting to win, with frontier models kept for the exceptions. The 90% saving, about a tenth of the old Anthropic and OpenAI bill, excludes cloud, staff and monitoring costs. The blended cost is about $0.30 per million tokens.

Revolut swaps Claude for GLM: who is affected

Roughly 2,000 Revolut software engineers, the whole engineering team, now use self-hosted GLM as the default for daily coding work. Claude Code and Codex remain the exception for the hardest work, so Anthropic and OpenAI no longer hold the daily default. Google stays in the workflow because Revolut rolled Gemini out first and still routes image tasks to Gemini. Engineers who saved rationed Claude Code tokens for hard tasks now have a flat self-hosted bill, and the per-engineer token limits are gone.

Z.ai provides the open weights, and Revolut does not use Z.ai's hosted service. Abitbol acknowledged the "stigma" of using a Chinese open-source model at a regulated bank. Revolut runs only those open weights on its own infrastructure in the US, with coding agents in a sandbox, approval rules, and a human in the loop. Smaller companies without Revolut's scale and infrastructure team may not get the same result from the 14-second to 1.7-second tuning on the same Nvidia B200 GPUs.

Revolut swaps Claude for GLM: what to watch

The 90% saving excludes cloud, staff and monitoring costs, and it depends on heavy tuning. Out-of-the-box inference was "terrible", Abitbol said, until time to first token fell from about 14 seconds to 1.7 seconds on the same GPUs, about 56 Nvidia B200s, at around $0.30 per million tokens. The 60% shorter coding time and 70% shorter lead time, from picking up a task to merging it, come from a comparison that does not isolate the model change from other causes, such as removing token limits.

The figures are from Abitbol's talk "Shipping like a startup, regulated like a bank" at the AI Engineer conference in New York, and they spread in early November after Trending Topics wrote them up. The live setup is the default for the whole engineering team, not as a pilot: GLM for daily coding, Gemini for images, and Claude Code plus Codex for the hardest work. For routine, high-volume work, "good enough and cheap" is starting to win, with frontier models kept for the exceptions.