Revolut runs GLM on its own GPUs for everyday work: how it did it and how to use GLM in Claude Code

Revolut, the digital bank with 80 million customers, no longer uses Anthropic and OpenAI models for its engineers’ everyday work. It uses GLM, an open-weight model from the Chinese company Z.ai, running on dedicated GPUs. Its AI bill is one tenth of what it used to pay Anthropic and OpenAI. David Abitbol, Ops & Infrastructure Leader at Revolut, shared this at AI Engineer New York on October 13, 2026.

What did Revolut announce at AI Engineer New York?

In the talk "Shipping like a startup, regulated like a bank", Revolut explained how it moved from paid models to an open-weight model hosted on infrastructure the bank chooses. The key points:

• 80 million customers and about 2,000 engineers involved;
• AI spend down to one tenth of what it paid Anthropic and OpenAI;
• about 56 B200 GPUs are enough to serve 2,000 engineers;
• coding time improved by 60% and lead time improved by 70%, according to Revolut’s internal data.

Frontier models haven’t disappeared completely: Revolut still uses Codex and Claude Code where they are needed. Everyday work, though, runs on GLM.

Why did Revolut drop Anthropic and OpenAI for everyday work?

It all starts with cost. With Claude Code and Codex, each Revolut engineer was spending between $1,000 and $10,000 a month on tokens, without a matching return.

Revolut set a spending cap for each engineer. The cap had a side effect: engineers were afraid of running out of budget. So they started saving their tokens for the hardest tasks and using AI far less for everything else.

What does open weight mean, and why did Revolut’s spend collapse?

An open-weight model is a model whose weights anyone can download and run on their own hardware. You don’t pay per request: you only pay for the servers it runs on. GLM is the open-weight model from Z.ai that Revolut picked.

Revolut deployed GLM on dedicated GPUs. From that point on, the bank stopped paying per token and started paying for infrastructure. The cost becomes fixed and predictable, and it no longer grows with every request engineers make.

In the talk, Abitbol says the bill is one tenth of what Revolut spent with Anthropic and OpenAI, with a blended cost of around $0.30 per million tokens. These figures reflect Revolut’s own setup and usage volume: they are not a price anyone hosting a model in-house can expect.

What changed for Revolut’s engineers?

With no budget to watch, Revolut’s engineers started using AI all day, for anything: code, email, documentation, research.

Revolut measured the impact on a sample of 2,000 engineers. According to the chart shown in the talk, with the arrival of GLM 5.1:

• coding time improved by 60%;
• lead time, the time until a feature is merged and ready to be deployed, improved by 70%.

These are Revolut’s internal figures, and they don’t isolate the model’s effect from other causes.

How do you use GLM in Claude Code without your own servers?

You can use GLM inside Claude Code too, without owning any servers. Instead of hosting the model, you use it through Z.ai’s GLM Coding Plan: Claude Code stays the same, but requests go to Z.ai’s servers.

Before you do, weigh the privacy trade-off. Your code and prompts go through Z.ai’s servers: Z.ai’s privacy policy says the service is generally provided from Singapore and that data is generally processed there. Read the terms and conditions, or have Claude read them with a prompt like this:

Read Z.ai’s privacy policy: what do they do with my data?

There are three steps.

1. On the Z.ai website, activate the GLM Coding Plan and create an API key.
2. In Claude Code, paste the link to Z.ai’s official Claude Code guide and ask Claude Code to configure itself. The prompt I use is: https://docs.z.ai/devpack/tool/claude configure yourself to use GLM 5.3. Claude Code asks for the API key you created in step 1: paste it when asked. Claude Code updates ~/.claude/settings.json on its own.
3. Close and reopen Claude Code.

From then on, Claude Code looks exactly the same, but GLM 5.3 does the work instead of Anthropic’s models. Z.ai’s guide points Claude Code to https://api.z.ai/api/anthropic and uses the Z.ai key instead of the Anthropic one.

When you need Opus for a hard task, just ask Claude Code: "For this hard task, switch back to Opus". Claude Code changes the settings again and you’re back on Anthropic’s model.

I’ve already covered a similar approach: letting a cheaper model work under Claude Code to save tokens. If you’d rather run a model in-house like Revolut, I’ve listed the open-weight coding models by hardware tier.

What does the Revolut case mean for anyone using Claude Code every day?

The Revolut case shows that per-token pricing decides how much people use AI. A spending cap protects the budget, but it pushes people to use AI only for hard tasks.

If you often hit Claude Code’s usage limit, you can try GLM for routine work and keep Opus for the tasks where it makes a difference. Weigh the privacy trade-off first, then decide how much of your projects you want to hand over to it.