Revolut Swaps Claude Code for Chinese GLM, Cuts Coding Costs 90%

Revolut has made one of the largest AI model vendor migrations in fintech history, shifting roughly 2,000 engineers from Anthropic's Claude Code and OpenAI's Codex to a self‑hosted deployment of GLM 5.3. The model comes from Z.ai, a Beijing‑based AI lab, and runs on Revolut's own infrastructure in the United States rather than in China.

David Abitbol, who leads AI operations and infrastructure at the $115 billion fintech, presented the migration at the AI Engineer conference in New York, describing it as a calculated bet on open‑weight models.,² the move reflects a broader internal conviction that "the bottleneck for engineering is not necessarily coding" but rather the non‑coding overhead that surrounds it.

The numbers driving the switch are stark. Revolut's AI coding costs fell to roughly a tenth of their previous spend after the migration, with token pricing now hovering around $0.30 per million tokens, ³ and corroborated by.¹

The entire operation runs on 56 Nvidia B200 GPUs serving all 2,000 engineers simultaneously. For a fintech of Revolut's scale — 80 million customers across 45 markets and reportedly targeting a $200 billion IPO — per‑engineer SaaS licensing for AI coding assistants had become a meaningful line item.

The migration wasn't plug‑and‑play. Out‑of-box inference on GLM 5.3 was, in Abitbol's own words shared at AI Engineer NYC, "terrible" — time to first token clocked in at roughly 14 seconds. After extensive tuning, Revolut brought that down to 1.6 seconds, a nearly 9x improvement, ¹ and.²

The gains extended well beyond latency. Revolut reported coding time down by roughly 60% and lead time — the full cycle from ticket to merged code — down by approximately 70%. The context window is capped at 256K tokens, a deliberate engineering choice. Abitbol reportedly described long‑context as "a scam" for most coding use cases,.²

A recurring question around the migration has been data sovereignty. GLM 5.3 is an open‑weight model from a Beijing‑based lab, but Revolut runs public model weights entirely on its own infrastructure in the United States, not in China. Abitbol acknowledged what he called "a lot of stigma" around using Chinese open‑weight models at an EU‑regulated financial institution,.³

The architecture means Revolut's code never leaves its own GPUs. The model weights are publicly available and run locally — there is no API call to Z.ai servers, no data exfiltration to Beijing, and no dependency on a Chinese cloud provider. For a regulated bank operating across 45 markets, that distinction is everything.

GLM 5.3 is not a full replacement for the previous toolchain. The model handles only text‑based coding tasks. For image‑related work, Revolut routes requests to Google's Gemini instead,.² Claude Code and Codex remain available to engineers but are now used sparingly rather than as the default.

This multi‑model routing architecture — GLM for text coding, Gemini for images, Claude and Codex as fallbacks — reflects a maturing approach to AI tooling in large engineering organizations. Rather than betting on a single vendor, Revolut has built a tiered system where cost, capability, and latency determine which model handles each request.

The GLM migration didn't happen in a vacuum. Revolut's AI coding journey went through several distinct phases, ² and.¹

The engineering team first experimented with Google's Gemini for AI‑assisted coding, which delivered roughly a 20% reduction in coding time — promising but not transformative. Next came a deployment of Claude Code with per‑engineer budgets, an attempt to give every developer access to a top‑tier coding model. That approach proved costly while delivering only modest incremental gains over the Gemini baseline. GLM 5.3 was the third attempt, and the one that stuck: open‑weight, self‑hosted, and tuned aggressively until the performance matched the economics.

Abitbol used his AI Engineer conference talk to make a broader point about engineering productivity. The real time sinks for Revolut's engineers are not in writing code — they're in "the inboxes, Jira, and Confluence" overhead that surrounds development work, ² and.³

An AI coding assistant, however fast or cheap, only addresses the narrow slice of an engineer's day spent actually typing code. Planning, coordination, documentation, and review consume the bulk of engineering hours. The 60% coding‑time reduction is real, but it applies to the fraction of work that was code‑generation in the first place. Abitbol's framing suggests Revolut sees the GLM migration not as an endpoint but as one piece of a larger effort to reduce non‑coding friction across the engineering org.