Revolut turns to self-hosted AI to cut coding costs tenfold over Claude

Embraces open-source models

British fintech firm Revolut says it has reduced the cost of running AI coding tools by about 90% after switching to self-hosted AI models, moving away from expensive commercial services such as Anthropic's Claude.

The company said the move had also helped improve coding time by 60% and reduced software delivery lead times by 70%, following the rollout of AI coding tools to about 2,000 engineers.

Addressing the AI Engineer conference in New York last month, Revolut's operations and infrastructure lead, David Abitbol, explained how the company had switched to using self-hosted AI models as its primary option in under a year.

Revolut initially experimented with Google's Gemini in early 2025 before adopting Claude Code in early 2026.

However, the company became concerned about the expense of commercial AI coding services, with some engineers spending between $1,000 and $10,000 a month.

Abitbol said the company initially introduced strict spending limits when it rolled out the tools.

But the restrictions limited the number of tokens engineers could use to process requests and generate AI responses.

The company subsequently began exploring self-hosted alternatives, adopting GLM-5.1 before switching to GLM-5.3, developed by Chinese AI company Z.ai.

Revolut said hosting GLM-5.1 cost about $0.30 per million input and output tokens, roughly a tenth of the cost of using services such as Anthropic's Claude Code or OpenAI's Codex.

The lower costs changed how engineers used AI, allowing them to delegate more routine work without worrying about exhausting their token allowances.

Abitbol said some of the biggest productivity improvements did not come directly from writing code faster, but from using AI to handle tasks such as debugging and managing tickets on project management platforms Jira and Confluence.

56 Nvidia GPUs power the AI infrastructure

Running the models at scale required significant computing resources.

GLM-5.1 has approximately 754 billion parameters, although around 40 billion are activated for each token generated.

To support GLM-5.3, Revolut deployed 56 Nvidia B200 graphics processing units (GPUs), providing the computing power needed to serve its engineers.

However, Abitbol said simply adding more GPUs was not the main factor behind the performance improvements.

Instead, the company focused on optimising the models' settings, introducing NVFP4 quantisation and separating the processing stages responsible for understanding prompts and generating responses.

It also adjusted context lengths, which determine how much information an AI model can consider during a task.

"Long context is a scam. Even though you can serve one million tokens, 256 [thousand] is really what you need," he said.

Following these changes, Revolut achieved a time to first token of less than three seconds for context windows of about 256,000 tokens, alongside generation speeds of approximately 200 tokens per second.

Abitbol said response times were particularly important for keeping engineers satisfied with the tools.

GPT and Claude remain part of the system

Despite the shift towards self-hosted models, Revolut has not abandoned commercial AI services altogether.

GLM is a text-only model, making it less suitable for tasks involving visual interfaces and other image-based work.

To address this, the company introduced a model router that directs different requests to the most appropriate AI system.

While GLM handles code generation, vision queries can be routed to models such as Google's Gemini or OpenAI's GPT.

Revolut believes the approach offers a more cost-effective way of using AI without sacrificing access to more specialised capabilities.

"Open source is now a priority at Revolut, and our economics are favourable," Abitbol said.