Revolut moves 2,000 engineers from Claude Code to GLM

Revolut has moved its roughly 2,000 software engineers from Claude Code and Codex to GLM, an open-weight model from Chinese lab Z.ai that it runs on its own infrastructure, and says the bill is now about a tenth of what it paid Anthropic and OpenAI, Trending Topics reported on November 3. The saving rests on a blended 30 cents per million tokens across 56 Nvidia B200 GPUs, which by our arithmetic only averages out that low at hundreds of billions of tokens a month. A StartupHub.ai analysis adds that the 90% leaves out cloud, staff and monitoring costs.

David Abitbol, who oversees AI operations and infrastructure at Revolut, laid out the numbers in a talk titled “Shipping like a startup, regulated like a bank” at the AI Engineer conference in New York. The full video has been on AI Engineer’s YouTube channel since October 29. Claude Code and Codex stay available. Abitbol called the latest Anthropic and OpenAI models excellent, but engineers now use them sparingly, as the exception.

How 56 GPUs serve 2,000 engineers

Revolut started on GLM-5.1 and has since moved to GLM-5.3, The Register reports. It serves the model on 56 B200s, the same GPU count as seven eight-GPU DGX B200 systems. Abitbol did not say whether Revolut bought the chips or rents them. He credited the biggest gains to tuning, not to extra GPUs: the model’s hyperparameters, a switch to NVFP4, Nvidia’s 4-bit number format, disaggregated prefill and decode, where reading the prompt and generating tokens run on separate groups of GPUs, and tuned context lengths.

Out of the box, inference was “terrible,” he said. Trending Topics puts the result of the tuning at time to first token falling from more than 10 seconds to 1.6 seconds, even as demand rose 76%. A summary that spread on Hacker News gives about 14 seconds down to 1.7 seconds on the same GPUs. Abitbol’s target is under 3 seconds. Above 5, “my Slack inbox starts to blow up.” Output runs at around 200 tokens a second.

Context is capped at 256,000 tokens. “Long context is a scam,” Abitbol said. “Even though you can serve one million tokens, 256 [thousand] is really what you need.” GLM handles only text, so a router sends image tasks to Gemini.

Why the token caps went

Gemini came first, in early 2025. By Abitbol’s account it cut coding time by 20% and lead time, the stretch from picking up a task to merging it, by 10%. Claude Code followed in early 2026 with strict per-engineer budgets. “We were highly skeptical about the cost of AI. We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI,” he said. The Register reports that the capped tokens are what pushed the company to look at self-hosting.

GLM has run on Revolut’s own infrastructure since the spring. In a sample of about 2,000 engineers, coding time fell 60% and lead time 70%. Abitbol reads the bigger drop in lead time as a sign that “the bottleneck for engineering is not necessarily coding,” since engineers lose time searching inboxes, Jira and Confluence. These are Revolut’s own figures, and the Hacker News summary notes that the comparison does not separate the model change from other causes, such as the end of token limits. With a flat bill, engineers no longer ration usage near the end of a quarter or on demanding tasks.

Chinese weights, US servers

Abitbol acknowledged “a lot of stigma in using Chinese open source model” at a regulated bank with more than 80 million customers in 45 markets. Revolut takes only the published weights and runs them on infrastructure in the United States. It does not go through Z.ai’s inference endpoint in China. Data is meant to be processed only in jurisdictions with equivalent data protection, such as the US, the EU, Britain and Israel, and providers must keep zero data retention by default. Revolut also tells regulators that its coding agents do not act autonomously. They run in a sandbox whose approval rules set which commands are allowed, and “they’re always with a human in the loop,” Abitbol said.