London-based fintech Revolut says it cut its AI coding costs by about 90% in under a year by moving roughly 2,000 engineers from Claude Code and OpenAI's Codex to a Chinese open-weight model that it runs on its own servers. The account, given by Revolut's AI operations and infrastructure lead David Abitbol at a New York conference and reported by The Register on 3 November 2026, is a rare, concrete case of a large company swapping out the leading commercial coding tools.
What Revolut did
According to The Register and other coverage, Revolut rolled out Gemini to engineers in early 2025 and tried Claude Code in early 2026, but kept strict budget limits because some engineers were spending between $1,000 and $10,000 a month without clear returns. It began self-hosting Z.ai's GLM-5.1 in April 2026 and has since moved to GLM-5.3. The model runs on 56 Nvidia B200 chips, equal to seven eight-chip DGX systems. Revolut reports a cost of about $0.30 per million tokens, roughly a tenth of what it paid before. Abitbol said that the changes made the engineers' use of AI no longer a question of rationing expensive tokens.
The results Revolut reports
- About a 90% fall in coding inference costs.
- A 60% improvement in coding time and a 70% improvement in lead time, which measures the period from picking up a feature to merging it. These are Revolut's own measurements.
- Throughput of up to around 800 tokens a second for engineers who sign in after 5 pm, when demand is lower.
Not a clean swap
The Register notes the model did not work well out of the box: the team had to cut the time to first response from 14 seconds to 1.7 seconds before full rollout, mainly through tuning rather than adding hardware. GLM is also text only, so Revolut uses a router that sends image and interface work to vision-capable models such as Gemini or GPT Luna while keeping code generation on GLM. Engineers can still use Claude Code and Codex when a task needs them, so the frontier tools have become the exception rather than the default.
What the number leaves out
The reported 90% figure is for inference costs. The coverage reviewed does not say how Revolut counts the upfront cost of the GPUs or the engineering time to tune and run the system, and a company without Revolut's scale and in-house skills would face both. Self-hosting only pays off when usage is high enough to keep the hardware busy. It also means the company takes on security, upkeep and model updates itself.
Why it matters
The story fits a pattern seen this week: AI coding is getting cheaper fast, through price cuts from the big labs, discounts to win large customers and strong open-weight models that can be run in-house. For startups and companies in Africa watching their cloud budgets, the lesson is not that everyone should buy 56 GPUs. It is that the cheapest sensible option now depends on volume, and that testing an open-weight model against a commercial one on your own tasks can be worth it.
Anthropic, the maker of Claude Code, is the company behind this assistant. This article reports Revolut's account and The Register's coverage, and it does not include a response from Anthropic.
Sources: The Register (3 November 2026); Cybernews; Crypto Briefing; Korben; ShiftDelete; Emergent.
