“Revolut cut its AI coding costs by 90% after switching from OpenAI and Anthropic to a Chinese AI model (GLM-5.3 by Z.ai)." Accompanying image shows crossed-out OpenAI and Anthropic logos over a Revolut building.”

The claim traces to a single origin: a conference talk by David Abitbol, who leads AI operations and infrastructure at Revolut, at AI Engineer NYC on October 13, 2026. The conference page carries a summary and a 202-passage timestamped transcript. Abitbol describes self-hosting GLM models as a way to give engineers much more freedom to use AI without unpredictable bills, with open models becoming Revolut's default for this workload and frontier services kept available for exceptions, and he reports using 56 B200 GPUs to serve GLM 5.3 to roughly 2,000 engineers. The transcript passages include the line placing the start of the programme at "April 26th, GLM 5.1" and a later line stating the current setup "is to host GLM 5.3." On cost, the quoted sentence from the talk is "Self-hosting GLM brings our blended cost of tokens to 30 cents per million. Our invoice is now one-tenth of what we were spending on Anthropic and OpenAI," said David Abitbol, who leads AI operations at Revolut. Revolut had been concerned about commercial AI costs, with some engineers spending $1,000 to $10,000 monthly without clear ROI, so it imposed strict usage budgets, and then looked for ways to give engineers unrestricted access without comparable commercial inference costs. In April 2026, Revolut began self-hosting open-source models, settling on the GLM family for coding workloads. One report adds a specific prior cap: before the change, Revolut set a $300 monthly Claude Code allowance for each engineer, and Claude Code and Codex are now mainly reserved for more complex coding tasks. On productivity, the conference page describes the measurement and its limits directly: Abitbol opens with two measures from a sample of about 2,000 engineers, coding time measured in hours and lead time from picking up a feature to having it merged and ready to deploy measured in days, and reports that the move to GLM 5.1 coincided with roughly 60% better coding time and 70% better lead time on Revolut's chart, while the internal comparison does not isolate the model change from other causes. An earlier company-wide Gemini rollout cut coding time by 20 percent and lead time by 10 percent, and Claude Code followed with strict per-engineer budget limits, where the benefit was small at first while costs were high. Abitbol's own interpretation is that when tokens feel scarce, engineers save them for a difficult coding task instead of using them across the surrounding work, and that engineering delays include the inbox, Jira, Confluence and other work around the code. Revolut did not drop the US providers. Revolut has not entirely replaced commercial models; its internally hosted GLM models handle most coding tasks. Image-related tasks still get routed to Gemini or similar hosted models, so the self-hosted path covers text and code generation only. GLM-5.3 is a real model. Z.ai's documentation describes GLM-5.3 as its latest flagship model, using the same base model as GLM-5.2 with all improvements driven by post-training. It is a large-scale reasoning model built for complex software engineering and long-horizon agent tasks, supporting a 1M-token context window. Weights are public but the flagship license is not MIT: Z.ai published the 753B-total-parameter GLM-5.3 weights on Hugging Face on August 28, 2026 under a bespoke licence tagged glm-5.3, while GLM-5.3-Flash shipped two days earlier under the plain MIT License.