Technology news and discussion on TalkFuse.
British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted
Revolut says it cut the cost of AI coding by about 90 per cent in under a year by making self-hosted models its default, rather than relying mainly on commercial tools such as Claude. The British fintech also reported faster coding and shorter delivery times across roughly 2,000 engineers, with cheaper access allowing AI to take on routine work beyond writing code.
The company moved from Gemini and Claude Code to GLM models developed by China’s Z.ai, and now serves GLM-5.3 on 56 Nvidia B200 GPUs. Revolut says GLM-5.1 tokens cost about $0.30 per million, around a tenth of the price of OpenAI’s Codex or Anthropic’s Claude Code. It tuned the system to deliver responses in under three seconds and around 200 tokens a second; other models remain in use for tasks such as working with images.
In short
- Revolut says self-hosted AI cut coding inference costs by about 90 per cent.
- Around 2,000 engineers use the tools, with coding time down 60 per cent.
- Cheap tokens let AI handle debugging and project tickets more freely.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Revolut’s case is that self-hosted models can make AI assistance affordable enough to use across routine engineering work, while giving the company more control over speed, availability and deployment. Its reported cost savings and faster delivery suggest that capable, lower-cost models may offer better value for many coding tasks than relying mainly on premium commercial tools.
The case against
The case for continuing to rely on commercial tools is that headline token prices do not capture the full costs of running, tuning and maintaining self-hosted systems, or the risks of depending on a narrower set of models. Commercial providers may offer stronger performance on some tasks, less operational burden and clearer support, so a mixed approach could be more reliable and cost-effective overall.
Explained simply
Revolut is a British financial technology company that offers banking and payment services through an app. Like other large software firms, it employs engineers to build and maintain its products, and has been exploring how artificial intelligence can help with that work.
AI coding tools can suggest or produce computer code from instructions, but they vary in price, speed and capabilities. Claude is a commercial AI model made by Anthropic, while GLM models are developed by China’s Z.ai; companies can access AI through outside providers or run models on their own computing systems.
The choice matters because AI use across a large engineering team can add up to substantial costs, while cheaper tools may make it practical to use AI for more routine tasks. Revolut’s approach also illustrates how businesses weigh the expense of running their own computing infrastructure against buying access to commercial services.