# Revolut Put 2,000 Engineers on GLM. Price Your Swap.

> Revolut says its self-hosted GLM costs about a tenth of its Anthropic and OpenAI bill. Price that swap before your renewal, then watch your code review queue.

- URL: https://www.iqsource.ai/en/blog/revolut-2000-engineers-glm-price-the-swap-before-renewal/
- Published: 2026-11-12
- Author: Ricardo Argüello, IQ Source

Revolut says it moved about 2,000 engineers to a self-hosted open model for coding, and that the invoice is now roughly a tenth of what it paid Anthropic and OpenAI. David Abitbol, who runs AI operations there, gave the numbers in a talk at the AI Engineer conference in New York in October. They are Revolut's own, from one talk, with no outside audit.

I am not telling you to copy a bank with 2,000 engineers. I am telling you to put a price on the swap before your next renewal. Revolut just handed every buyer of AI coding tools a reference number, and your vendor can read the same coverage you did.

## What the talk actually says

The [AI Engineer page for the talk](https://ai.engineer/talks/Ok6QdZHCcI8-revolut-cut-ai-coding-costs-90-self) and a [detailed write-up by SparseNotes](https://sparsenotes.com/posts/2026/10/revolut-self-hosted-open-models-david-abitbol/) agree on the core. GLM, an open-weight model from Z.ai, is the default coding model for about 2,000 engineers. It runs on 56 Nvidia B200 GPUs at a blended $0.30 per million tokens, with cache hits free. Claude Code, Codex and Gemini stay available for tasks where they are clearly better.

Abitbol reported coding time down 60% and lead time down 70%. [Trending Topics](https://www.trendingtopics.eu/revolut-glm-open-weight-coding-costs/) carries the same figures. Nobody gives a baseline for either, and the AI Engineer page notes that the trend does not separate GLM from other changes at Revolut.

Two cautions on the cost claim. The "90%" compares two invoices, without matched usage or periods. And [Silicon Report](https://www.siliconreport.com/revolut-says-self-hosted-ai-could-cut-coding-costs-by-roughly-90) says it cannot tell whether $0.30 covers hardware, power and staff. Read it as a signal about order of magnitude, not as a quote you can paste into a budget.

Where it runs also depends on which source you read. SparseNotes says Together AI dedicated inference, and the AI Engineer page names no provider. I would not repeat either version as settled.

## Price the swap before the renewal

A closed vendor sets its price against your alternatives, and until now your alternative was another closed vendor. A bank just published a third option with a number attached. I made a similar argument about [what you are paying a closed vendor for](/en/blog/muse-spark-open-price-ceiling-closed-ai-apis/), and about [writing a routing rule](/en/blog/grok-bot-model-routing-buying-policy/) that decides which model gets which task.

Revolut itself did not leave. Frontier models are the exception list, and a router sends image work elsewhere because GLM is text-only. The swap is a routing decision.

So price it. Take the last invoice and divide it by the engineers who use the tool. Abitbol said some of his engineers had been spending somewhere between 1,000 and 10,000 a month (the page does not state the currency). Find your own number. Then quote an open-weight model at your token volume on rented GPUs, add the engineering hours to run it, and test it on twenty tasks from your own repo. If the gap is small, you renew without drama. If it is large, you negotiate with a number.

## When tokens stop being rationed, review becomes the limit

The detail I would keep from the talk is about rationing, not price. Before the switch, Revolut gave each engineer a token budget. In SparseNotes' account, Abitbol said the gains under those limits were small, and that removing the budgets mattered more than which model ran.

Lead time also improved more than coding time. SparseNotes reports he read this as evidence that coding was not the main bottleneck, and that engineers started pointing agents at email, Jira and Confluence. Coding agents there run with a human in the loop.

The talk does not discuss code review. That is my inference, so weigh it as one. Cost used to be a brake, something I wrote about in [Cost Was Your AI Guardrail](/en/blog/cost-was-your-ai-guardrail/). Take the brake off and engineers run more agents, in parallel, all day. Every change still lands in front of a person who has to read it and decide whether to merge. GPUs scale. Reviewers do not.

I covered the cost side in [Code Is Not Cheap](/en/blog/code-is-not-cheap/): generating code was never the expensive part. A cheaper model makes that more visible.

## What I would do on Monday

Three numbers, one afternoon. Last invoice divided by active users. Pull requests merged per week. Median hours a pull request waits for its first review.

The third one is rarely on hand. If it is already two days, a model that doubles your output will not double your delivery. It will double the queue.

If you want a second pair of eyes on those numbers before a renewal call, that is the kind of pricing we do at IQ Source. [Write to us](/en/contact/) with your invoice and your pull request counts, and we will tell you which side of the table has the leverage.

## Frequently asked questions

### How much does Revolut's self-hosted GLM cost compared with Claude Code and Codex?

David Abitbol of Revolut said the invoice is about a tenth of what the bank previously paid Anthropic and OpenAI, at a blended $0.30 per million tokens on 56 Nvidia B200 GPUs. Those are Revolut's own figures from one conference talk. Coverage does not say whether hardware, power and staff are included.

### Did Revolut stop using Claude Code and Codex after switching to GLM?

No. Per the talk, GLM became the default coding model for about 2,000 engineers, and Claude Code, Codex and Gemini stay available for tasks where they are clearly better. Image work goes to a separate vision-capable model through a router, because GLM is text-only.

### How do I price a switch from a closed AI coding vendor to an open model before renewal?

Take your last invoice and divide it by the engineers who use the tool. Then estimate monthly token volume, quote GPU hosting or a dedicated inference provider for an open-weight model at that volume, add engineering time to run it, and test the model on your own code. The gap is your negotiating number.

### Why does cheaper AI coding make code review the new bottleneck?

When per-engineer token budgets disappear, engineers run more agents and produce more changes. Every change still needs a person to read and merge it, and reviewer hours do not scale with GPUs. Revolut's talk kept coding agents under human supervision, and its lead time gain was larger than its coding time gain.