{
  "id": 18016623,
  "title": "British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted",
  "url": "https://urgent.news/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-03T20:07:29.000Z",
  "source": {
    "name": "The Register",
    "slug": "the-register",
    "url": "https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345"
  },
  "original_language": "en",
  "account": "British fintech company Revolut has dramatically reduced its AI coding inference costs by 90% by switching to self-hosted models, according to its operations and infrastructure lead David Abitbol. Speaking at the AI Engineer conference in New York, Abitbol revealed that the company saw a 60% improvement in coding time and a 70% improvement in lead time after deploying AI coding tools to roughly 2,000 engineers. Initially using GLM-5.1, Revolut later adopted GLM-5.3, which is served by 56 Nvidia B200 GPUs.\n\nPrior to this transition, Revolut had been experimenting with AI, deploying Gemini in early 2025 and Claude Code in early 2026. The company was initially cautious about AI costs, with engineers spending up to $10,000 a month without significant ROI. To mitigate this, Revolut implemented strict budget limits, restricting engineers' token usage to a limited number. This led to the exploration of self-hosting models, which proved to be significantly more cost-effective at $0.30 per million input/output tokens.\n\nAbitbol emphasized that while GLM-5.1 is a powerful model requiring substantial GPU memory, the company achieved the most significant performance gains through hyperparameter tuning, NVFP4 quantization, disaggregated prefill and decode, and optimizing context lengths. They found that a 256,000 token context window was sufficient for their needs, debunking the notion that longer context lengths are inherently beneficial. Revolut achieved a sub-3-second time to first token and a generation speed of 200 tokens per second, which they deemed ideal for their software engineers' workflow.\n\nThe Chinese open weights model GLM-5.1 is not a replacement for models like GPT or Claude, as it is text-only and better suited for code generation. Revolut uses a model router to direct vision queries to models like Gemini or GPT Luna while handling code generation with GLM-5.3. While open source models are now a priority for Revolut, they acknowledge potential privacy and security concerns associated with running these models on-premises or on regional cloud resources.",
  "summary": "Turns out cheap tokens get engineers doing things that save time, like debugging and tackling Jira tickets",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Business Insider",
        "title": "Claude's new AI report card tool gives companies a closer look at costs — and flags 'inefficiencies'",
        "url": "https://urgent.news/2026/11/02/claudes-new-ai-report-card-tool-gives-companies-a-closer-look-at",
        "published": "2026-11-02T21:15:27.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}