Revolut Cuts AI Coding Costs 10x by Switching 2,000 Engineers to GLM 5.3 2026

Revolut has made GLM 5.3, the flagship AI model from Chinese company Z.ai, the default coding assistant for its approximately 2,000 engineers, according to David Abitbol, Revolut's head of operations and AI infrastructure. The move marks a significant shift away from frontier tools like Anthropic's Claude Code and OpenAI's Codex, which are now reserved only for the most complex tasks.

Speaking at Nvidia's GTC conference in Berlin, Abitbol revealed that Revolut now pays roughly one-tenth of what comparable frontier models cost. For a fintech serving over 80 million customers worldwide, that translates into massive savings without sacrificing engineering productivity.

Why Revolut Switched from Claude Code and Codex to GLM 5.3

Until recently, Revolut's engineers relied heavily on Claude Code and Codex for day-to-day programming assistance. These frontier models offered strong performance but came with premium price tags. As AI usage scaled across thousands of developers, costs ballooned.

GLM 5.3 changed the calculus. It is an open-weight model, meaning its parameters are publicly available, and Revolut runs it on its own servers. This self-hosted approach gives the company greater control over data privacy, latency, and—most importantly—spend.

Abitbol noted that frontier tools remain available but are now treated as a reserved option with a dedicated budget. Engineers can still use them for tasks that genuinely require frontier-level reasoning, but routine coding work flows through GLM 5.3 by default.

How Much Cheaper Is GLM 5.3 Compared to Frontier Models?

According to Abitbol, Revolut's AI bill is now about one-tenth of what it would be with comparable frontier models. That 90% reduction is not a marginal improvement—it is a structural change in unit economics.

For a company with 2,000 engineers, even small per-request savings compound quickly. Multiply that across code completions, refactoring suggestions, test generation, and documentation, and the annual difference becomes enormous.

The table below summarizes the key differences between Revolut's previous setup and its current approach.

What This Means for the AI Coding Tool Market

Revolut's decision is a bellwether. If a heavily regulated fintech with 80 million customers can trust an open-weight model for everyday engineering work, other enterprises will take notice. The stigma around open-weight models is fading fast.

Two years ago, self-hosted models lagged far behind frontier APIs. Today, the most capable open-weight models can handle tasks that once seemed impossible—and they do so at a fraction of the cost while keeping data local. That combination is hard to ignore.

Anthropic and OpenAI still lead on the hardest reasoning problems, and Revolut keeps their tools on hand for exactly that reason. But the default layer of AI assistance is shifting toward cheaper, controllable alternatives.

Key Takeaways for Engineering Leaders

• Cost reduction is real: Revolut reports paying about one-tenth of frontier model costs after switching to GLM 5.3.
• Open-weight models are viable: A fintech with strict compliance needs runs GLM on its own servers for daily coding work.
• Frontier models still matter: Claude Code and Codex remain available for complex tasks, but with a dedicated budget.
• Scale amplifies savings: With 2,000 engineers, even small per-request savings add up to a massive annual difference.
• The market is shifting: More enterprises are experimenting with self-hosted models as performance improves.

What to Watch Next

Revolut's experiment will be closely watched by other large engineering organizations. If productivity holds steady while costs drop by 90%, expect a wave of similar migrations. The key question is whether GLM 5.3 can match frontier models on the long tail of complex debugging and architecture tasks. The background to this story is set out in Revolut Outage: App and Card Payments Down for Millions. A related thread runs through Revolut Free ChatGPT Go: Up to a Year for Paid Plan Users. Earlier coverage in Revolut Becomes Spain's Fourth Largest Bank by Customers shows how the argument developed.

Abitbol's comments suggest Revolut has already found the right balance: use open-weight models for the bulk of work, and reserve frontier tools for the edge cases. That hybrid approach may become the standard playbook for enterprises adopting AI coding assistants. A closer look at the same subject appears in Z.ai GLM-5.3 Coding Gains Boost Cybersecurity Skills Fast. Anyone following the wider picture can start with Revolut Data Breach: Hackers Demand $3 Million Ransom After Phishing Attack. The same theme surfaces in Revolut Account Freeze Puts £37,000 House Deposit at Risk, which takes a different angle.

For now, Revolut's 2,000 engineers are voting with their keystrokes—and the bill is ten times smaller. Context on how this played out before sits in Business: latest articles. A companion piece, Barcelona Becomes Fintech Hub for Europe's Biggest Neobanks, approaches the same subject from another angle.