Artificial intelligence programming assistants save developers time, but their bills can quickly skyrocket. The British neobank Revolut chose a radical path to control it: largely replacing commercial services like Claude Code with a open model hosted on its own servers. Announced result: a drop of approximately 90% of the cost in less than a year. The testimony, reported on November 3, 2026 by the website The Register, sheds light on a debate that concerns all companies.
Exploding bills
David Abitbol, head of operations and infrastructure at Revolut, detailed the approach at the AI Engineer conference in New York in October 2026. The company has deployed AI programming tools to around 2,000 engineers. She first used Gemini from Google, in early 2025, then Claude Code from Anthropic, in early 2026.
The problem quickly became apparent: some engineers were spending $1,000 per monthsometimes up to $10,000, with no commensurate gain. Revolut therefore imposed strict spending limits, before thoroughly reviewing its strategy. With thousands of users, the bill for a service billed per usage can indeed reach considerable amounts.
Choosing an open model
The neobank opted for GLMa family of open models developed by the Chinese company Z.ai, whose weights, i.e. the parameters that define the model, can be downloaded and installed freely. She changed the version used over the months, with reports of the intervention diverging on version numbers.
The model runs on 56 Nvidia B200 GPUsthe equivalent of seven DGX B200 servers. Recent GLM models are based on a so-called expert mixture architecture: only a small part of their hundreds of billions of parameters is activated for each word generated, which reduces the necessary computing power.
Spectacular figures
According to Revolut, the hosted model costs approximately $0.30 per million tokensthe tokens being the text fragments processed by the model. That’s about a tenth of the cost of using OpenAI’s Codex or Anthropic’s Claude Code.
| Indicator | Figure announced by Revolut |
|---|---|
| Falling cost of programming AI | Around 90% in less than a year |
| Cost of hosted model | Around $0.30 per million tokens |
| Engineers equipped | About 2,000 |
| Material | 56 Nvidia B200 GPUs |
| Generation speed | Around 200 tokens per second, more during off-peak hours |
| Save on programming time | 60% |
| Savings on delivery time | 70% |
Revolut did not specify how these productivity gains were measured, and these figures have not been independently verified.
Adjustment rather than power
The most instructive is the method. According to David Abitbol, the performance gains came mainly from fine adjustments rather than purchasing additional hardware: adjusting parameters, moving to a more compact representation of the model’s numbers, called quantization, separating the steps of reading the request and generating the response, and limiting the context length.
On this last point, the manager believes that 256,000 context tokens are actually sufficient for the needs, even if the model can handle a million of them. Reducing the working memory allocated to each user allows more engineers to be served with the same hardware. According to a conference report published by StartupHub, these tweaks helped improve performance by 88% while demand increased by 76%, without adding GPUs. The manager also recognized that the model installed as is, without adjustment, gave poor results: in-house hosting requires real know-how.
Debugging rather than writing code
Another lesson: according to Revolut, the biggest gains did not come from writing code, but from debuggingi.e. finding and correcting errors, and processing tickets in the Jira and Confluence project management tools. An observation that is consistent with other feedback: AI especially helps to understand existing code and accelerate repetitive tasks.
Business models not abandoned
Revolut has not given up on commercial services. The GLM model only processes text and is therefore not suitable for working on GUIs. A router directs requests: code remains handled by the hosted model, while visual tasks are assigned to commercial models like Gemini.
David Abitbol also addressed the question of trust in a Chinese model. He acknowledges privacy and security concerns, especially when accessing these models through online services outside China, and believes hosting on its own servers or in a regional cloud addresses them.
Conditions difficult to reproduce
The StartupHub report highlights an important limitation: Revolut’s results are based on a fleet of 56 high-end graphics processors, on continuous adjustment and on the ability of a bank to set the rules for using its tools itself and to distribute the load according to working hours. There is no guarantee that another organization would achieve the same savings under different conditions.
What businesses can learn from it
Revolut’s experience cannot be transposed as is to all organizations. It requires significant material investment and teams capable of operating large models. But it provides several useful benchmarks:
- measure spending per user from the start and set ceilings;
- identify truly profitable useslike debugging, before generalizing;
- combine models: a hosted model for volume, commercial services for specialized tasks;
- finely tune deployment rather than multiplying equipment;
- host the models locally or in a controlled region for sensitive data.
For European companies, the option of an open, internally hosted model also exists with players on the continent, such as Mistral AI, which recently published the weights of Mistral Large 4. The choice will depend on each team’s on-task performance, licensing conditions and security requirements.
Pressure on code assistant prices
The case of Revolut illustrates an underlying trend: the performance gap between open models and commercial models is narrowing, while the prices of commercial services remain high for intensive use. Suppliers are also reviewing their formulas, such as OpenAI with its usage reservation common to ChatGPT and Codex. For large companies that equip thousands of developers, the question of the cost of AI becomes as strategic as that of its quality.

