FAIND – Flash AI News DailyFAIND – Flash AI News Daily ← All the news

Page language: Italiano English Français Deutsch

FAIND › The point

The point

Revolut and GLM: what actually happened, and what the headlines got wrong

In brief

What was said

According to reports from The Register and Trending Topics, published on 3 November, Revolut moved from Gemini (early 2025) to Claude Code (early 2026) and then, because of cost, to tighter and tighter token budgets. It then chose to host an open-weight model itself: first GLM-5.1, then GLM-5.3, on about 56 Nvidia B200 GPUs in US infrastructure, not on Z.ai's own service.

The stated numbers: about $0.30 per million tokens, around a tenth of the previous spend; coding time down about 60% and time from picking up a task to merging it down about 70%; time to first token down from over 10 seconds to about 1.6. The model handles text only, so image tasks go to other models such as Gemini.

Five claims going around, and why they fail

"Revolut dropped Claude Code and Codex." It didn't. They stopped being the default and remain available as the exception for the hardest work. This is routing, not a breakup.

"AI coding is now 90% cheaper." The comparison is per token, and Revolut makes it. It excludes cloud, staff and monitoring. Fifty-six B200s and the team that tunes them are a real cost, and no total has been published. A tenth of the API bill is not a tenth of the cost.

"A Chinese model beat the American ones." Nothing in the account compares quality. Revolut picked a model that is good enough for daily work and cheap to serve, and it keeps frontier models for the hard cases.

"The new model made engineers 60% faster." The data does not separate the model's effect from other changes, in particular the removal of token limits. According to the account, the biggest gains came from debugging and tickets, not from writing code. Our reading is that the likelier lesson is that people work better when nobody has to ration the tool.

"Anyone can do this." Revolut has thousands of engineers and an infrastructure team. Before tuning, response speed was, in the speaker's own word, "terrible". A small company renting GPUs risks paying more than a subscription.

What it means

The default is what counts. When a tool is cheap enough that nobody counts tokens, people use it all day. That is where the stated gains come from.

Frontier models become the escalation tier. A cheap model for routine work, an expensive one for the hard cases. Expect more large companies to set things up this way, not necessarily with GLM.

For regulated companies, where the model runs matters more than where it was built. Hosting it in a controlled environment is what makes a Chinese-origin model acceptable to a bank. Revolut openly acknowledged the stigma.

For small teams, the lesson is method, not hardware. Route by task, cap spend per tool, measure before and after.

What we don't know

There is no independent measure of code quality or bug rates before and after the switch, and no total cost including hardware and staff. The GPU price estimates in the coverage are the journalists', not Revolut's. No regulator comments have been reported.

Frequently asked questions

Did Revolut stop using Claude Code and Codex?

No. They are no longer the default but remain available as the exception for the hardest work.

Is the 90% saving confirmed?

It is Revolut's own claim and refers to cost per token. It excludes cloud, staff and monitoring, and nobody outside the company has verified it.

Sources