Revolut has moved its roughly 2,000 software engineers off Anthropic’s Claude Code and OpenAI’s Codex as their default coding tools, and says the switch has cut its AI coding inference bill by about 90%. The engineers now work by default with GLM, an open-weight model from Chinese AI lab Z.ai, which Revolut hosts itself. Claude Code and Codex remain available, but only as the exception for the hardest work.
The figures come from David Abitbol, who runs AI operations and infrastructure at Revolut, in a talk titled ‘Shipping like a startup, regulated like a bank’ at the AI Engineer conference in New York on 13 October 2026, reported by The Register, Cybernews and Dealroom on 3-4 November. Revolut’s path to the switch ran through Google’s Gemini in early 2025 and Claude Code in early 2026. Abitbol said the early Claude Code rollout was costly for a modest gain, with some engineers spending $1,000 or even $10,000 a
month, which led Revolut to impose strict per-engineer budget limits. Engineers then rationed tokens, saving AI help for the hardest coding tasks.In April 2026 Revolut began hosting GLM 5.1 itself, and it now runs GLM 5.3 on about 56 Nvidia B200 GPUs. At a blended cost of around $0.30 per million tokens, self-hosting costs roughly a tenth of what Revolut used to pay Anthropic and OpenAI for comparable token use. Abitbol reported roughly 60% shorter coding time and 70% shorter lead time, from picking up a task to merging it, after the rollout to about 2,000 engineers. 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. The 90% figure also excludes cloud, staff and monitoring costs, and Revolut has not published its method or invoices, so no third party has been able to redo the maths.The self-hosting detail matters for the privacy question. Revolut says it uses only the open weights and runs them on its own infrastructure in the United States, not through Z.ai’s hosted service, and keeps a human in the loop on coding tasks. Image work is routed elsewhere, to Gemini, because GLM is text-only. Abitbol acknowledged that Chinese open-weight models can raise privacy and security concerns, but said the real risk lies in using APIs hosted in China. The wider point is about cost control rather than a claim that open models beat frontier models outright. Revolut’s billing is now flat, Abitbol said, with no bill shock at the end of the month, and the frontier is the exception. For other companies the $0.30 figure will not transfer directly: it reflects Revolut’s deployment, utilisation and heavy tuning, and out-of-the-box inference needed work before it performed well. What is not claimed: GLM has not replaced Claude or GPT entirely, the savings are not independently verified, and the productivity gains are Revolut’s own measurements. What is confirmed by multiple independent reports of the same talk is the switch itself, the default-to-exception change for Claude Code and Codex, and the scale: about 2,000 engineers on 56 B200 GPUs.FAQ: What did Revolut switch to? GLM, an open-weight model from Z.ai, self-hosted by Revolut, currently version 5.3. Did it drop Claude Code completely? No — Claude Code and Codex remain available for the hardest work. Is the 90% saving verified? No — it is Revolut’s claim from a conference talk, excluding cloud, staff and monitoring costs, and not independently audited.
Related reading: Oracle Cancels $7B Tencent AI Chip Lease | Apple Warns AI Agents Could Override User Instructions | DeepSeek Plans to Pre-Train V5 as China Chases AI Self-Sufficiency
Sources
- The Register coverage via Cybernews
- Dealroom: Revolut swaps Claude Code for a Chinese open-weight model, cutting its AI costs
- Korben: Revolut ditches Claude Code for Chinese AI, cuts costs tenfold
