Revolut cuts its AI costs by 90% by abandoning Claude
The British fintech says it has reduced its inference costs for code tenfold by switching to self-hosted models. It also reports time savings for 2,000 engineers but keeps a router for visual tasks.
Revolut, a British fintech, claims to have reduced its AI inference costs related to code by about 90% in less than a year by switching to self-hosted models rather than relying mainly on commercial services like Claude. The company says this choice has also improved the daily work of around 2,000 engineers, with a 60% gain in coding time and 70% in processing delays. This topic matters because it affects the bill, productivity, and the choice of AI tools in business.
When tokens cost less, engineers really use them
David Abitbol, head of operations and infrastructure at Revolut, explained these figures at the AI Engineer conference in New York last month [according to the source]. Before this shift, the company had tested Gemini in early 2025 and then Claude Code in early 2026, while imposing strict budget caps. The result: some engineers saw bills of 1,000 dollars per month, sometimes 10,000 dollars per month, with a return deemed too low.
The switch to GLM-5.1, then GLM-5.3, changed the equation. Revolut indicates it now serves this model with 56 Nvidia B200 GPUs and bills usage around 0.30 dollars per million input/output tokens. At this level, the cost is about ten times lower than OpenAI Codex or Claude Code, according to the company. It’s not just a matter of price: when tokens become cheap, teams stop counting each request as a lost office euro.
The real gain is not just from code
According to Abitbol, the main benefit is not limited to writing code. The biggest gains come from tasks that nibble away time without writing an extra line, such as debugging, Jira tickets, or Confluence. With more accessible tokens, Revolut says its teams use AI more broadly, without having to reserve their requests for the most complex cases or the end of the quarter.
The company does not present GLM as a universal solution. The Chinese model used by Revolut is textual, so poorly suited for visual interface tasks. For these cases, the fintech says it uses a model router to direct image or interface requests to Gemini or GPT Luna, while keeping code generation on GLM. This hybrid approach avoids relying on a single provider for the entire chain.
An optimized architecture before simply bigger
The deployment did not consist of stacking GPUs hoping everything would work out. Revolut says it adjusted its hyperparameters, adopted NVFP4 quantization, implemented decoupling of pre-filling and decoding, then fine-tuned context lengths. The GLM-5.1 model is massive, with about 754 billion parameters in total and 40 billion activated per generated token, requiring about 380 GB of GPU memory for weights, plus 6 to 26 GB per user depending on precision.
In this context, Revolut considers a one-million-token context window of little practical interest. Abitbol summed up the position bluntly: "Long context is a scam. You really need to be mindful of that," adding that a 256,000-token window is really sufficient for their use. With these settings, the company says it achieves a first-token access time under three seconds on windows of about 256,000 tokens, as well as a throughput of about 200 tokens per second. In the evening, after 5 p.m., this figure would even rise to around 800 tokens per second, according to the executive.
What Revolut really wants to avoid
The fintech’s message is clear: enterprise AI must not become an unpredictable budget line. "Open source is now a priority at Revolut, and our economics are favorable. The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," said Abitbol. The company acknowledges privacy and security issues around Chinese open models but considers the main risk comes from access via APIs hosted in China. In local execution or on regional clouds, Revolut believes the risk level remains controlled.
Ultimately, the challenge goes beyond Revolut. If a large fintech can industrialize the use of open models to reduce costs, other groups could follow, especially where code, internal support, and document automation dominate. There remains a clear limit: the economics are favorable only if one accepts moving from "all with the same provider" to a more technical, finer management, and a bit less comfortable for purchasing teams.
By the numbers
- 90% — announced reduction in AI inference costs in less than a year at Revolut.
- 2,000 engineers — internal scope concerned by the deployment, according to Revolut.
- 56 Nvidia B200 GPUs — infrastructure used to serve GLM-5.3, according to the company.
- 0.30 dollars — announced cost per million input/output tokens with self-hosting.
- 754 billion parameters — approximate total size of GLM-5.1, according to the source.