Skip to content
Articles / eigen-ai-model8 min read

Revolut dropped Claude:what you can learn from it

Revolut replaced Claude with its own open AI model. Costs fell by about 90%, and its developers got faster. What a small business can take from that.

Published
By
Luuk
Reading time
8 min read
About
eigen-ai-model · self-hosting

Revolut, the digital bank with over 80 million customers, has taken Claude away from its developers. They now simply run their own open AI model. As a result, the bill per use is about 90% lower. And what I actually find more interesting: the developers work faster with it too.

Revolut first gave its people Claude, with strict budgets on top. If you keep watching the meter, you use AI sparingly, and then it doesn't deliver much. With your own model you pay the same every month. That's when people dare to use it for everything.

You don't have to be a bank to do something with this. For a company of 10 or 50 people there are a few lessons in it too, I think:

  • A fixed bill works better than a meter. People only really start using AI when they're not afraid of the invoice.
  • The most expensive model doesn't have to be the default. For most work, a good open model is enough.
  • If you run the model yourself, your data doesn't go to America or China.

At STAIBL we run our own AI model too, although we started for a different reason. For us it was mainly about our customers' data, which had to stay in Europe. But the benefit Revolut is so happy about now, we see as well: our customers pay a fixed price per month, however many questions they ask. Further down you can read how we go about it.

01

What happened at Revolut

Revolut first tried Google's Gemini and then switched to Anthropic's Claude. Because they were worried about the costs, everyone got a tight budget. "We were highly skeptical about the cost of AI," infrastructure lead David Abitbol said at a conference1.

That was disappointing. Some heavy users spent thousands of dollars a month on tokens, but the team got barely any measurably faster. So since April the bank has been running GLM, an open model from the Chinese company Z.ai, on its own servers: 56 Nvidia B200 cards. About 2,000 developers now work with it2.

What Revolut reports

per million tokens

$0.30

About a tenth of what they paid before.
faster coding

60%

faster from idea to delivery

70%

to the first word

1.6 s

That was over 10 seconds, before tuning.

"The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," says Abitbol. They still use Claude and Codex, but only for what the open model can't do. Images, for example, go to Gemini.

One caveat: these are figures from a presentation, and nobody has checked them independently. The $0.30 doesn't include the servers themselves, the staff or the monitoring. It didn't happen by itself either, because the model only performed well after they had tuned it considerably.

02

Why this isn't just about banks

Revolut has 56 graphics cards that each cost tens of thousands of euros. As a small business you're obviously not going to copy that. But the problems Revolut ran into are ones you'll have at any company that gets serious about AI.

The first is the meter. If you use Claude, ChatGPT or Gemini through an API, you pay for every word that goes in and out. It's a bit like eating out and paying per bite: you order carefully. Employees then experiment less, while that's exactly where the gains are. At Revolut the biggest jump also came from the work around coding, such as sorting email and processing meetings. You only automate that once nobody is counting tokens anymore.

On top of that, you don't know what it will cost next month. The vendor changes the price or the limit, and your budget no longer adds up. TechRadar even calls that unpredictability a bigger problem than the high price.

And your data goes along. Everything you ask an American model ends up with an American company. For a bank that's a compliance question, but for you too, as soon as there are customer details, quotes or HR matters in it.

With your own model you don't have those problems. You pay a fixed price per month and the data stays on a server you know. That's exactly why we at STAIBL run our own AI model.

Work it out yourself

When does your own AI model pay off?

Set how many people at your company work with AI and what that costs per person per month in subscriptions or API use. Where the lines cross, your own model becomes cheaper.

People working with AI15
160
Now per person per month€40
€0€300
Subscription or APIOwn model
060 people
Your own model pays off from
32people
Now per month€600
Own model per month€1,250

With 15 people, a subscription or API is still cheaper. The break-even point is at 32.

Not sure, or does privacy matter more to you than price? Then we'll work it out together.

Book a call

We assume €1,250 per month for your own model: a GPU server of around €1,000 plus a few hours of management. It's an estimate. Whether your data may leave the building counts just as much in the choice.

03

What you can take from it

What I take from it: choose the model that fits each job, and make the cheap model the default. In practice it comes down to this.

  1. First list what you use AI for now: answering email, summarising, quotes, customer questions, code. Usually most of it is just routine work.
  2. Let an open model do that routine work. You can download Gemma, Qwen, Mistral and GLM for free, and for most everyday work they're good enough. Do test a few on your own tasks. What scores well in other people's rankings won't necessarily work best for you.
  3. Only use the top model when you have to. A tricky analysis or code where a lot depends on it can go to Claude or GPT, as long as that's a deliberate choice.
  4. Choose where the model runs: on your own server, or with a European party that runs it for you. With an open model from China or America, that location is what matters most. If it runs on a European server, nothing goes back to the maker.
  5. Count the whole bill. Revolut says so itself: the price per token doesn't include hardware, power and management. So put what you spend per month now next to the fixed price of your own or a rented server, and include someone to look after it.
  6. Expect to tune. At Revolut the model was slow out of the box, and only after tuning did the first word arrive within 2 seconds. Honestly, that's half the work.
04

How we do it

We actually do the same as Revolut, just at small-business scale. And with one extra rule: everything stays in Europe.

We have one server of our own with one graphics card. It runs an open model for all our customer tools, such as the chat on websites, processing conversations and searching documents. Customers pay a fixed price per month, so whether someone asks ten questions or ten thousand makes no difference to the bill.

Every tool that uses the model is identifiable for us. Per tool we see how often it calls the model and how long it has to wait. That way we see in time when the server is starting to fill up.

When the server is busy, a question simply waits in the queue for a moment. We don't switch to another service then. Only if it's really broken does a European party take over, and so never an American or Chinese one.

That's where we go further than Revolut, because the bank also runs the model in the US and Israel. For our customers "processing in the EU" is a promise, and we don't break it, not even for one question.

05

When you shouldn't do it

Your own model isn't the smart choice for everyone. I'd stick with a subscription or API if:

  • You only use AI now and then

    If you spend a few tens of euros a month on tokens, you'll never earn back a server that costs hundreds of euros a month.

  • You need the very best model for almost everything

    Open models are good, but for the heaviest thinking the American top models are still ahead. That's why Revolut keeps them in reserve too.

  • Nobody can look after the server

    Updates, tuning and monitoring have to be done by someone, yourself or outsourced. Install and forget doesn't work here.

  • You mainly process images or video

    Open models are less far along there, and Revolut sends images to Gemini as well.

Frequently asked questions

Why did Revolut stop using Claude?

Because of the costs, and because the budgets held back use. With strict limits on Claude, the developers got barely any faster. With their own open model, without a meter, they worked 60% faster on code.

Which model does Revolut use now?

GLM, an open model from the Chinese company Z.ai, on its own servers with 56 Nvidia B200 cards. The bank only uses Claude and Codex as the exception.

Isn't a Chinese open model unsafe?

That depends on where it runs. An open model you put on your own server doesn't send any data back to the maker. If you use the maker's own API, your data does go to that company.

Can a small business do the same as Revolut?

At a smaller scale, yes. One server with one good graphics card is enough for a model of around 30 billion parameters and several tools at once. You don't need 56 cards.

Is your own AI model cheaper than Claude or ChatGPT?

If you use it a lot, yes. The price per question drops sharply, but you pay a fixed price for the server and its management. With little use, a subscription is cheaper.

Should I stop using Claude or ChatGPT altogether?

No, Revolut still uses them for the heaviest work. It's just that its own model is the default there, and the expensive model the exception.

Rather not build it yourself?

Revolut has its own infrastructure team, and you probably don't. With Eigenwijs we run an AI model of your own for you on a European server. You pay a fixed price per month and your data stays in the EU.

Talk to us about what your own AI could do for your business.

Sources
  1. 1TechRadar · techradar.com
  2. 2StartupHub · startuphub.ai