One of the talks doing the rounds this week is Revolut’s, from the AI Engineer conference in New York. As Computing and The Register report it, Revolut moved its roughly 2,000 engineers off Claude Code and Codex as the default, and onto open-weight GLM models from Z.ai that it hosts itself (on 56 Nvidia B200 GPUs, which at least should keep the server room toasty this winter 🙂). The headline numbers: inference at around $0.30 per million tokens, about a tenth of what they were paying before, plus claims of coding time down 60% and delivery lead times down 70%. Frontier models are still there for the hard cases.

A 90% cut on what is, nowadays, one of the bigger lines on an engineering budget will get any CFO’s attention. So, should you do the same? My answer is a very engineer-ish “it depends”, but let me explain, because I think the more interesting lesson isn’t about which model they picked.

Good Enough Is the Real Benchmark

Here’s the thing: most of what an AI coding agent does all day is not hard. It’s renaming things, writing tests, wiring an endpoint, fixing the lint, reading a stack trace. Using the best frontier model for all of that is a bit like hiring a Formula 1 driver to do the school run. Sure, they’ll get there, and in style, but you’re paying for a skill you don’t need on that particular journey.

What Revolut seems to have done is accept that “good enough for 90% of tasks” at a tenth of the price beats “the best for 100% of tasks” at full price, as long as you keep an escape hatch to the best model for the remaining 10%. That’s a routing decision, and to be honest, it’s the one most teams should be making whether they self-host or not. You can get a decent chunk of the same benefit by sending routine work to a cheaper hosted model.

Self-Hosting Needs a Team

Now, before anyone orders a rack of GPUs on the strength of a conference talk: running models yourself is a real commitment. Somebody has to own capacity planning, upgrades (they’ve already moved from one GLM version to the next since April), latency, security patching and evaluation. Revolut has around 2,000 engineers to spread that cost across; at that scale the GPUs and the platform team pay for themselves. With 40 engineers, I’d argue the maths looks very different, and the hosted options will quite likely stay cheaper once you count the people.

There’s also due diligence that a regulated business has to do on any model it runs, wherever it comes from: what it was trained on, how it behaves on your code, what your auditors will ask. Self-hosting helps with some of that (your code never leaves your infrastructure, which a bank will like a lot) and adds homework elsewhere.

What I’d Actually Take From It

I wrote recently about being able to switch AI providers by Friday. Revolut is the best real-world argument for that I’ve seen: they could only make this move because their tooling let them swap the model underneath without retraining 2,000 people on a new workflow. If your prompts, agent instructions and tool plumbing are welded to one vendor’s SDK, you can’t take advantage of a price drop like this, however good it is.

So, if I were a CTO looking at this on Monday morning, I’d do three things. First of all, measure which tasks your agents actually spend tokens on; I suspect most teams have no idea. Secondly, build a small evaluation suite on your own codebase, so “is this cheaper model good enough?” becomes a test you run in an afternoon rather than a debate. And last but not least, make the model a configuration value, not an architecture decision.

As it turns out, I suspect the most valuable thing Revolut built is the freedom to change its mind about models every few months, more so than the GPU cluster. Given how fast this market moves, that freedom is worth far more than any single discount.