Revolut says it cut AI coding costs by 90% with an open model
The British fintech swapped pay-per-use AI for a self-hosted open model for 2,000 engineers. What the math shows and what it leaves out.

Revolut says it cut its AI costs for programming by about 90% after making a self-hosted open model its default instead of pay-per-use services. The British fintech serves about 2,000 engineers with 56 Nvidia GPUs and models from the Chinese company Z.ai, according to a report by The Register published on Tuesday (November 3), about a talk by David Abitbol, from the company’s operations team, at the AI Engineer conference. AI Chat Daily confirms the core figures.
How to read that number
The data comes from Revolut itself, and The Register doesn’t report the baseline or the method behind the 90%. The cost cited, US$ 0.30 per million tokens, doesn’t include cloud, staff and monitoring.
The company also hasn’t abandoned paid models. The open model only works with text, so requests with images go to other providers. Abitbol summed up the logic by saying the cost is fixed and the frontier model has become the exception.
Before, according to him, some engineers were spending US$ 1,000 to US$ 10,000 a month with little return. After the switch, the biggest gain showed up outside code writing, in debugging and internal support tickets. Revolut reports a 60% improvement in coding time and 70% in delivery time.
What this changes for your company
1. Compare the total cost, not the token price. Include infrastructure, people and monitoring. Few companies have 2,000 engineers to spread that structure across. For most, the way forward may be an open model contracted from a provider.
2. Decide which task deserves the most expensive model. This part applies at any size. Classify usage and reserve the frontier model for what needs it. The bill has gone over budget at American companies, and this is a concrete brake.
3. Treat an open model as a supplier. Abitbol says the risk with Chinese models lies in using them through APIs operated in China, not in running them on your own infrastructure or a regional cloud. Even so, a recent test with an Alibaba model shows why it’s worth limiting what it can access and logging what it does.
In 29 years of programming, I’ve seen a lot of savings vanish once the cost of operating comes in. It’s worth doing the math before changing the architecture.
Originally published on IWA7 Insights, IWA7 Digital’s blog, which is now part of Escovando Bytes.


