Revolut cut its AI coding bill 90% by swapping Claude for self hosted GLM

One of Europe's biggest fintech companies has quietly stopped treating Claude and Codex as the default for its engineers. Revolut now runs a Chinese open weights model, GLM 5.3, on its own GPUs and hands it to about 2,000 engineers as the standard coding assistant. The frontier models from Anthropic and OpenAI are still available, but only as the exception. The headline number is blunt: Revolut says its AI coding bill is now about one tenth of what it used to pay Anthropic and OpenAI.

The details come from a talk that David Abitbol, who leads AI operations and infrastructure at Revolut, gave on the main stage of AI Engineer NYC on October 13, 2026. The Register picked up the story on November 3, and it has been travelling through developer circles since. This article first lays out what Revolut actually said, what it measured and what it did not. After that, clearly separated, comes djcroman's opinion on what it means for the fight between closed frontier labs and open models.

What Revolut did

Revolut describes itself as a digital bank in Europe with 80 million customers in 45 markets, and Abitbol opened his talk with that context for a mostly American audience. The company started using AI in engineering in January 2025 with Google's Gemini, which Abitbol credits with modest gains: roughly 10% better lead time and 20% better coding time, mostly because engineers no longer had to dig through Stack Overflow and documentation.

Claude Code came next. Revolut experimented with it in 2025 and rolled it out at scale in early 2026. According to Abitbol, the economics were the problem. "We were highly skeptical about the cost of AI. We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI," he said in the talk. So Revolut launched Claude Code with strict budget limits.

Those limits created a second problem. When tokens are scarce, engineers save them for the hardest coding task of the week and do everything else by hand. Revolut wanted the opposite: AI used freely on the boring work around code, such as debugging, Jira tickets and Confluence pages. That pushed the team toward open models it could host itself, where the marginal cost of one more prompt is close to zero once the GPUs are paid for.

Around April 2026, Revolut began self hosting GLM 5.1 from the Chinese lab Z.ai. It later moved to GLM 5.3, which it now serves to roughly 2,000 engineers on 56 Nvidia B200 GPUs. In GPU count, that is the equivalent of seven eight GPU DGX B200 systems, as The Register notes. "Open source is now a priority at Revolut, and our economics are favorable. The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," Abitbol said.

The numbers, and what they do and do not prove

The cost claim is specific. Abitbol said the blended cost of a token on Revolut's own setup is about 30 cents per million, and that "our invoice is one tenth of what we were spending on Anthropic and OpenAI." The official talk summary on the conference site adds an important caveat: those figures describe Revolut's own deployment and utilization, not a portable price that any company will get by downloading the same weights. If your GPUs sit idle half the day, your cost per token goes up quickly.

The productivity claim needs even more care. Revolut tracks two measures across its sample of about 2,000 engineers: coding time in hours, and lead time in days, from the moment an engineer picks up a feature until it is merged and ready to deploy. After the move to GLM 5.1, Abitbol reported roughly 60% better coding time and 70% better lead time. The conference summary is explicit that this internal comparison does not isolate the model change from other causes. The biggest single change may simply have been removing the token caps, not the model itself.

Abitbol's own explanation points the same way. Lead time improved more than coding time, which he reads as proof that the bottleneck in engineering is not typing code. "You can save a lot of time not having to dig through your inbox or messing with Jira and Confluence," he said. "When you have unlimited tokens, you're free to do that. You no longer have to save your tokens for the end of the quarter or the most difficult task."

So the honest summary is this: Revolut got a much cheaper bill and, at the same time, much faster engineering metrics. It has not shown that GLM is a better coding model than Claude or Codex. It has shown that a good enough model with no meter running can change how a large engineering team works.

How they made a giant model fast enough

GLM is not a small model. The Register puts GLM 5.1 at roughly 754 billion total parameters, with about 40 billion active for each token, and estimates around 380 GB of GPU memory just to hold the weights. Serving that to 2,000 impatient engineers is an infrastructure project, not a weekend experiment.

Abitbol's target numbers are practical. Time to first token has to stay under three seconds. "As soon as you exceed 5 seconds, my Slack inbox starts to blow up," he said. Throughput matters less: around 200 tokens per second is fine, because engineers increasingly start several agent tasks in parallel and come back when they are done, instead of watching the reasoning stream. Engineers who log in after 5 pm, when the GPUs are quiet, can get close to 800 tokens per second.

The most useful part of the talk for other teams is how they got there without buying more hardware. According to the slide Abitbol showed, time to first token fell from 14.26 seconds during the migration to 1.68 seconds in the stable setup, an 88% improvement, while the load served rose by 76%. The GPU count stayed at 56 the whole time.

The levers he named were mostly configuration. First, he called long context "a scam": even if the model can handle one million tokens, Revolut found that 256,000 is what coding work really needs, and reserving memory for more steals room from other users. Second, on Blackwell GPUs he recommends FP4 quantization, which he says cuts the GPU memory footprint by about 40% compared with FP8 without a loss in precision he could measure. Third, even though GLM fits on four GPUs, running it across eight leaves more room for the KV cache, the memory that stores attention data for tokens already processed. Finally, he advised watching the cache hit rate, which for coding should be above 90%, and looking at speculative decoding and separating the prefill and decode stages.

There are limits. GLM is a text only model, so it cannot look at screenshots or UI mockups. Revolut handles that with a model router that sends image tasks to Gemini or OpenAI models and keeps code generation on GLM. Abitbol also mentioned a smaller vision capable GLM variant for the UI design team. All endpoints speak an OpenAI compatible API, so engineers can keep using harnesses like OpenCode or the Codex app.

The China question

Abitbol addressed the obvious objection head on. There is, he said, "a lot of stigma in using Chinese open source models." His answer is that the risk depends on the data path, not the model's passport. Revolut does not send prompts to an inference endpoint in China. It downloads the weights and runs them on infrastructure it picks, in jurisdictions covered by data adequacy rules such as the US, EU, UK and Israel. The conference summary frames it the same way: downloading weights and running them yourself is a different data path from calling a vendor's hosted service.

As an EU regulated bank, Revolut also needed to convince supervisors that these tools are not autonomous agents. Its answer is a sandbox with explicit rules about which commands an agent may run without approval. Abitbol's example: allow grep, but not gradle. The sandbox also protects against insiders misusing the tools and against prompt injection. Strong telemetry is the last piece, so the team can reconstruct what happened if something goes wrong.

The conference write up adds a fair warning: Revolut's view of its own regulatory position should not be read as a general ruling that any model, region or sandbox satisfies another company's obligations. Self hosting moves responsibility for cost control, security and observability onto your own team.

djcroman's take: Chinese open models will win enterprise coding on price, not on benchmarks

This section is opinion. Everything above is reported fact; what follows is how djcroman reads it.

For two years the debate about open versus closed models has been framed as a leaderboard race. Who has the best SWE bench score, who tops the arena this week. Revolut's talk suggests that framing misses where enterprise money actually goes. Most coding work inside a large company is not frontier work. It is reading logs, fixing flaky tests, updating tickets and refactoring code nobody wants to touch. For that work, a model that is slightly behind the frontier but costs a tenth as much, with no meter running, is not a compromise. It is the better product.

And right now the most capable open weights models for coding largely come from Chinese labs such as Z.ai, DeepSeek and Qwen. Western frontier labs mostly keep their best weights closed and sell access by the token, while European companies like Revolut face strict rules about where their data goes. Self hosted Chinese weights solve both problems at once: the price is set by your own GPUs, and the data never leaves your chosen cloud. That is a structural advantage that does not depend on winning any benchmark.

There are real counterarguments, and they deserve a fair hearing. Revolut is not a typical company. It has the scale to keep 56 B200s busy and the engineers to tune its own serving stack. A 200 person startup will not get $0.30 per million tokens. Frontier models still win on the hardest tasks, which is why Revolut keeps Claude Code and Codex for exceptions. And the productivity numbers cannot be cleanly credited to GLM at all. There is also a political risk: if Western governments restrict Chinese weights the way they restrict Chinese hardware, a company that bet its developer workflow on them could have to switch overnight.

The money data still points one way. Menlo Ventures' 2026 State of Enterprise AI report found that open source models already handle 19% of tokens but only 4% of enterprise LLM spend. That gap is exactly what Revolut is exploiting. Once finance teams see a bank cut its AI coding bill by 90% while its delivery metrics improved, "default to the frontier API" stops being the safe choice and starts needing a justification. Our bet: within a year, the default coding model at many large companies will be an open weights model they host themselves, very often a Chinese one, with the frontier labs as the expensive specialist you call when the cheap one gets stuck.

What you can do now

• If you run engineering, measure where your AI tokens actually go. If most of them are routine work, price out a self hosted open model for that slice before renewing per seat or per token contracts.
• If you self host, follow Revolut's order of operations: tune context length, quantization, parallelism and caching before buying more GPUs.
• Treat data path and model origin as separate questions. Running downloaded weights in your own cloud is a different risk from calling an API in another country.
• Keep a router. Text only open models still need a fallback for vision tasks, and a frontier model for the hardest problems.

Sources

• The Register: British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self hosted (Nov 3, 2026)
• AI Engineer NYC 2026: Shipping like a startup, regulated like a bank, David Abitbol, with transcript (Oct 13, 2026)
• Talk video on YouTube: How Revolut cut AI coding costs 90% with self hosted open models
• AI Engineer NYC 2026 field guide
• Menlo Ventures: 2026 State of Enterprise AI