British financial technology company Revolut says making self-hosted AI models its default for coding has cut its inference bill (the cost of running AI models to process requests and generate responses) by roughly 90 per cent, allowing engineers to use the technology more freely while retaining commercial services for tasks that need them.
What Did Revolut Do?
Revolut began introducing AI tools in early 2025 and later expanded access to Claude Code, Anthropic’s coding assistant. However, the cost of processing requests and generating responses made the company cautious about how much engineers could use it.
Speaking at October’s AI Engineer conference in New York, David Abitbol, who leads AI operations at Revolut, described imposing strict budgets after seeing substantial spending without an equally clear return. Those restrictions controlled expenditure, but also encouraged engineers to reserve their allowances for difficult coding problems.
Revolut therefore began hosting open-weight models in April 2026, i.e., using AI systems whose developers make the learned settings available for others to download and run. This allows Revolut to run the models on infrastructure it chooses, rather than sending every request to the developer’s own service.
What Is Revolut Running?
Revolut started with GLM-5.1, a model from Chinese AI developer Z.ai, before upgrading to GLM-5.3. It runs the service for around 2,000 engineers using 56 NVIDIA B200 chips designed to handle demanding AI tasks.
Abitbol said the average cost was around US$0.30 per million tokens, the small chunks of text an AI model reads and produces. He explained, “our invoice is one-tenth of what we were spending on Anthropic and OpenAI.”
Those savings reflect how Revolut runs and uses its service, so other businesses should not expect the same reduction automatically. The figure also concerns the cost of running the AI, rather than the full cost of developing software, which still includes staff, maintenance and checking the results.
Why Did Cheaper Access Make A Difference?
With less need to ration requests, engineers could use AI for work they might previously have considered too routine to justify the expense. Abitbol highlighted time spent searching inboxes and working with Jira and Confluence, tools used to manage tasks and documentation.
That matters because writing code is only part of delivering a feature. For example, an engineer may also need to find an earlier decision, understand a reported fault or collect information before making a change. Reducing those delays can therefore help work progress even when the code itself takes little time to write.
Revolut reported improvements of around 60 per cent in coding time and 70 per cent in lead time, measured from picking up a feature to having it merged and ready for deployment. It should be noted that these were internal observations rather than a controlled comparison, so they do not really establish how much improvement came from the model change alone.
How Did It Keep The Service Responsive?
Making the service affordable also required work on how it ran. Revolut says it adjusted settings to suit engineers’ requests, reduced the memory needed to run the model and avoided reserving capacity for unnecessarily large amounts of background information.
Its presentation showed the time before a response began falling from 14.26 seconds to 1.68 seconds while requests per second increased, without adding GPUs. For engineers waiting for assistance, that responsiveness helped make the cheaper option practical.
Revolut has also kept access to Claude, Codex and other commercial services. Its main coding model handles text, while requests involving images can go elsewhere. Choosing a model according to the task lets the company control costs without expecting one system to do everything.
Could Other Businesses Follow?
Gartner’s October research forecasts that open-weight models will perform 90 per cent of enterprise agentic workflow steps. Its public summary highlights routing work between models, supporting the idea that routine operations need not always use the most expensive available service.
However, individual steps and completed tasks are different measures. A cheaper model may need more attempts or human corrections, so businesses should compare what they spend to obtain an acceptable result.
It's worth noting here that Revolut’s scale also matters. For example, demand from around 2,000 engineers can keep substantial infrastructure busy, whereas a smaller business could pay for capacity it rarely uses. Hosted services may remain better value where demand is limited or unpredictable.
What Does This Mean For Your Business?
For UK businesses reviewing AI spending, a useful first step is to identify which tasks account for the bill and what staff gain from them. That could reveal opportunities to use a cheaper model for routine work while retaining a more capable service for difficult requests.
Any comparison should include hosting, support, maintenance and the time spent checking results. It should also establish where information is processed and who can access it. Running a model on chosen infrastructure gives greater control over that arrangement, but does not automatically make its output reliable or its actions safe.
Revolut’s presentation described approval rules, restricted working environments and activity records as part of making the system workable. Its experience suggests that lower costs can encourage more useful experimentation, provided businesses can manage the resulting increase in use. The lasting saving will come from completing worthwhile work reliably, with enough oversight to catch mistakes before they become expensive.