Revolut switches to an open AI model for its developers
The online bank says it has cut its AI bill by ten by running GLM-5.1 on its own servers. It also claims 60% less coding time and 70% faster delivery times.
Revolut says it has scaled up its use of AI for its developers by switching to GLM-5.1, an open-weight model from the Chinese lab Z.ai run on its own servers. The European online bank claims to have reduced coding time by 60% and delivery times by 70%, while lowering its AI bill to a tenth of what it previously paid to Anthropic. The message is twofold: productivity gains exist, but the real issue also becomes infrastructure, data control, and the real cost of AI in business.
Une facture d’IA qui ne se paie plus à la requête
Revolut's switch to GLM-5.1 mainly tells a story of a business model change. Where services like Claude, Gemini, or other commercial models are often billed by token volume, the bank chose an open-weight model that it installs itself on dedicated machines. In other words, it now pays for servers rather than variable usage. The model is not free, however: David Abitbol, head of AI operations at Revolut, mentions 56 Nvidia B200 chips and an average cost of 30 cents per million tokens.
This choice is far from trivial for technical departments. Usage-based pricing can quickly escalate when teams massively adopt code assistants. Revolut says some engineers spent $1,000, then $10,000 per month on Claude Code, without satisfactory returns. By switching to internal capacity, the bank offers its engineers almost unlimited tokens — though the machine must still be sized as demand grows. Freedom of use has its price, but it is more predictable. And for a product team, a budget that stops melting like ice on a radiator often has a certain appeal.
Pourquoi GLM-5.1 a davantage compté que Gemini ou Claude Code ?
According to David Abitbol, Revolut first saw gains with Gemini in early 2025, with 20% less coding time and 10% faster delivery. Then the bank tested Claude Code, with little effect at the time. The real leap came in April 2026, when switching to GLM-5.1, a model usable without token limits according to the presentation given in New York in mid-October, then republished on YouTube on October 29.
The most telling progress was not the code itself, but delivery time — the time between taking charge of a feature and its deployment. Revolut draws a clear conclusion: code was not the main bottleneck. Once the budget constraint was lifted, engineers also entrusted AI with peripheral tasks, such as searching their email or internal tools. The open model thus serves not only to write faster; it also broadens what a team is willing to delegate.
La question des données reste sous surveillance
Revolut stresses a sensitive point: its data remain in protected environments compliant with European requirements. David Abitbol specified before his American audience that engineers do not use a service that sends requests to China; they simply download the model weights to install them on a US-based service. According to him, there is no problem as long as data remain in the European Union or in countries deemed equivalent in data protection.
The bank also says it runs these assistants in an isolated environment, with predefined operations and a human in control. This response reassures about data flow, but does not cover everything: risks related to the model itself, nor those of the code it generates, have not been detailed. In this context, Revolut plays the card of regulatory pragmatism. It does not give up American models: Claude Code, Codex, and Gemini remain available for some uses, even if open models have become a priority.
Ce que ce choix dit du marché de l’IA en entreprise
Revolut's decision matters beyond its own IT department. It confirms that an open-weight model can attract a regulated company when it promises control, predictable cost, and productivity. It also shows that enterprise AI is not just about raw performance. The debate now focuses on the bill, operational sovereignty, and how to keep control over data without restricting usage.
One important limit remains: Revolut does not detail the specific risks of the Chinese Z.ai model nor the possible security trade-offs accompanying its adoption. This silence is not uncommon in public demonstrations, but it leaves a gray area. For now, the bank mainly highlights a simple observation: when engineers no longer watch their token consumption at every click, they use AI more. And apparently, they do not use it only to code faster.
By the numbers
- 60% — reduction in coding time, announced by Revolut.
- 70% — reduction in delivery times claimed by Revolut.
- 56 Nvidia B200 chips — hardware capacity used to serve engineers.
- 2,000 engineers — measurement base cited by Revolut for its statistics.
- 30 cents per million tokens — average cost reported by David Abitbol.