---
id: "lib-2026-11-03-british-fintech-revolut-slashes-ai-coding-costs"
type: "library"
title: "British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted"
lang: "EN"
org: "The Register"
source_type: "Report"
date: "2026-11-03"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-03-british-fintech-revolut-slashes-ai-coding-costs/"
original_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"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "The Register's systems editor reports a conference talk by Revolut's ops and infrastructure lead with concrete serving details; a credible first-hand account, but every outcome number is self-reported and unaudited."
summary_basis: "full-text"
original_minutes: "5"
tags: ["enablement", "business transformation", "open weights", "self hosting"]
---

# British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted

Tier B (Useful, limited): The Register's systems editor reports a conference talk by Revolut's ops and infrastructure lead with concrete serving details; a credible first-hand account, but every outcome number is self-reported and unaudited.

Summary based on the full text.

## In short

A 3 November 2026 Register article by Tobias Mann on a talk by David Abitbol, ops and infrastructure lead at Revolut, at the AI Engineer conference in New York. Revolut moved its AI coding tools from commercial services such as Claude Code to self-hosted open-weight GLM models from Z.ai and says inference costs fell about 90%. Cheap tokens changed use: engineers stopped rationing and applied the models to debugging and ticket work, where Abitbol says the biggest gains came.

## Key numbers

- roughly 90 percent: cut in Revolut's AI coding inference costs in under a year after making self-hosted models the default
- $0.30: cost per million input/output tokens of self-hosted GLM 5.1, about a tenth of Codex or Claude Code
- 56: Nvidia B200 GPUs serving GLM-5.3 for roughly 2,000 engineers
- 60 percent / 70 percent: reported improvement in coding time and in lead time after rolling out AI coding tools
- sub 3-second: time to first token at about 256,000-token context, with about 200 tokens a second generation

## What it reports

- Gemini was deployed in early 2025 and Claude Code in early 2026, with strict budget limits after engineers spent $1,000 to $10,000 a month "with not so much ROI".
- Self-hosted GLM 5.1 at $0.30 per million tokens cost about a tenth of Codex or Claude Code; Revolut later moved to GLM-5.3.
- GLM 5.1 has roughly 754 billion parameters (40 billion active per token) and needs about 380 GB of GPU memory for weights plus 6-26 GB per user for context.
- GLM-5.3 runs on 56 Nvidia B200 GPUs. The main performance gains came from tuning, NVFP4 quantization, disaggregated prefill and decode, and context-length limits, not extra GPUs.
- Targets: under 3 seconds to first token at around 256,000 tokens of context, about 200 tokens a second; Abitbol: "Long context is a scam."
- Reported 60% improvement in coding time and 70% in lead time for roughly 2,000 engineers.
- GLM is text-only, so a router sends vision queries to Gemini or GPT Luna; frontier models are "the exception".

## How they got it

A journalist's write-up of a conference talk, with the reporter's own context on model size and hardware. No slides, baseline definitions or cost breakdown are included.

## Limits

- All outcome figures come from Revolut; how coding time and lead time were measured is not stated.
- The $0.30 per million token figure appears to cover inference only; staff, operations and whether the GPUs are owned or rented are not stated.
- No quality comparison between GLM and the commercial models on Revolut's own work is given.

## Why it matters for companies

At high volume, the cost of coding agents is driven by token pricing, and a self-hosted open-weight model can change the economics enough to remove rationing. The useful detail is operational: tuning and context limits mattered more than adding GPUs, and a router keeps frontier models for the cases that need them. Firms considering the same move should cost in the serving team, compare quality on their own tasks, and check data-residency rules for models run on their own or regional infrastructure.

Original: 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