Going headless like Revolut: what replacing Claude Code with an open model took
Revolut now runs most of its AI coding on an open-weight model it serves itself, for about a tenth of its former bill. A reconstruction of what it built, what the numbers mean for other companies, and the step that comes after.
Contents
The bill that started it
Revolut did not set out to make a statement about open models. It set out to control a cost. Some of its engineers were running up $1,000 to $10,000 a month each on AI coding tools, with little measurable return, and the company introduced strict per-engineer budgets1. David Abitbol, its operations and infrastructure lead, described the mood plainly: the team was highly sceptical about the cost of AI1.
The sequence matters, because it is the sequence most large engineering organisations are now in. Revolut deployed Gemini in early 2025 and Claude Code in early 20261. The Gemini rollout, which had access to internal data, cut lead time by about 10% and coding time by about 20%, mostly by replacing documentation lookups. Claude Code under tight budgets produced little measurable effect at first2. Budgets kept the bill in check, but they also kept engineers from using the tools freely.
What Revolut built
In April 2026 Revolut began moving its default coding model from commercial services to GLM-5.1, an open-weight model from the Chinese developer Z.ai, and later to GLM-5.321. It downloaded the weights and served them on dedicated hardware, so no request goes to Z.ai's own endpoints2. The deployment serves about 2,000 engineers1.
| Component | What Revolut reported |
|---|---|
| Default model | GLM-5.1 from April 2026, then GLM-5.3 (text only)1 |
| Model size | GLM-5.1: about 754 billion parameters, about 40 billion active per token1 |
| Hardware | 56 Nvidia B200 GPUs, the equivalent of seven eight-GPU systems1 |
| Interface | OpenAI-compatible endpoints, usable from OpenCode and Codex2 |
| Context limit | Capped at about 256,000 tokens; the full million was not needed1 |
| Latency targets | First token in under 3 seconds; complaints start beyond 5 seconds1 |
| Other models | A router sends image tasks to vision-capable models; frontier models stay for the hardest work12 |
| Controls | A sandbox that asks a person to approve non-allowlisted commands, zero data retention at the inference provider, OpenTelemetry logs2 |
The most transferable finding is about where the performance came from. On the same 56 GPUs, tuning the serving configuration raised throughput from about 3 to more than 7 requests a second and cut the time to first token from 14.26 to 1.68 seconds. Abitbol attributed about 80% of the gain to serving parameters rather than hardware2: 4-bit NVFP4 quantisation, separating the prefill and decode stages, tuned context lengths and sticky sessions that keep prompt caches warm12. For coding work he expects cache hit rates above 90%2.
What it changed
Revolut reports three results, all self-reported in the talk:
- Cost. AI coding inference costs fell by about 90% in under a year, at a blended $0.30 per million tokens, about a tenth of the Claude Code and Codex bill1.
- Speed of work. Coding time improved by about 60% and lead time by about 70% after the rollout to about 2,000 engineers1.
- Use. With a flat, predictable bill, engineers used agents for debugging, Jira and Confluence work, not only code generation. Abitbol said the biggest gains came from that work1.
His summary of the policy was direct: "Open source is now a priority at Revolut"1.
The arithmetic for everyone else
Revolut has not published what it pays for its GPUs. Notes from the talk say they run as dedicated inference at Together AI2, so that provider's public list price gives a reasonable outer bound. At Together AI's on-demand price for dedicated B200 inference, $8.99 per GPU per hour3, the calculation for a deployment of Revolut's size is as follows.
| Item | Calculation | Result |
|---|---|---|
| Hourly cost, 56 B200 GPUs | 56 × $8.99 | $503 an hour |
| Monthly cost, running all month | × 730 hours | about $367,500 |
| Per engineer, 2,000 engineers | ÷ 2,000 | about $184 a month |
| Same size on Together GPU Clusters, preemptible | 56 × $4.09 × 730 | about $167,200 a month |
These figures are our calculation from list prices, not Revolut's costs, and they exclude engineering time. They still show the shape of the decision. A flat cluster costs the same whether engineers send ten prompts or ten thousand, which is why usage rose once the budgets came off. Below a certain volume, per-token open-model endpoints are cheaper: Together AI lists GLM-5.3 at $1.40 per million input tokens, $0.26 for cached input and $4.40 per million output tokens3. The break-even point depends on volume and cache hit rate, which is why Revolut's first step was measurement.
What Revolut has not done yet
The published account describes a company that owns its inference. It does not describe a company that owns its intelligence. Nothing in the talk or the coverage mentions fine-tuning GLM on Revolut's code, tickets or documents2. The model knows what any copy of GLM-5.3 knows, plus whatever context is placed in the prompt at the moment of the request. The earlier Gemini rollout drew its gains from access to internal data2, which suggests where the next gain lies.
That next step is harder than serving a model, and the difficulty is mostly not in the training. It is in the data. Company knowledge sits in tickets, mail, documents, chat threads and system records, written for people, duplicated and often contradictory. Before any of it can train a model, it has to be found, cleaned, labelled and structured, and that work has to continue as the company operates. Our guide to training an AI model on company data sets out the stages.
This is the layer OpenLazarus works on. Its architects build a data lake that holds all of a company's data, owned by the company rather than its software vendors4. Its AI Router sends each request to a frontier model, an open model or a model trained for the customer, and it offers custom training on that data5. In effect it acts as the company's librarian: it puts the knowledge in order so that a model can learn from it. It is not a model itself; it is the means of producing one that belongs to the company. OpenLazarus has set out its own view of this next step in an article on Revolut's move.
The path, in five steps
- Measure before switching. Record tokens, cost and latency per team and per tool for at least a month. Revolut's case rested on knowing that a few engineers were spending thousands a month with little to show for it.
- Put a router in front of every model. Point agents and tools at one OpenAI-compatible endpoint you control. Swapping, testing or self-hosting a model then becomes a configuration change.
- Move the default, not everything. Choose an open model for the bulk of the work and keep frontier models for the tasks where they are measurably better, as Revolut did for images and the hardest problems.
- Tune serving before buying hardware. Set targets for time to first token and throughput, cap context at what people actually use and measure cache hits. Revolut got most of its gain this way.
- Organise the data, then train. Build an inventory of the company's knowledge, an evaluation set drawn from real tasks and a pipeline that keeps both current. Fine-tune an open model on it, route the work it does best to it, and keep measuring.
The end state is concrete. Most requests go to a model the company controls, at a cost that does not rise with every prompt. The same endpoint can serve a model that has learned the company's own systems, terms and decisions, and the frontier labs become one option behind the router rather than the foundation.
Questions buyers ask
What open-source model does Revolut use?
Revolut's default for AI coding is GLM from Z.ai. It started with GLM-5.1 in April 2026 and moved to GLM-5.3. Frontier models from Anthropic, OpenAI and Google remain available for the tasks where they are clearly better, and a router sends image work to vision-capable models.
How much did Revolut save by self-hosting?
Revolut reports that its AI coding inference bill fell by about 90%, to a blended $0.30 per million tokens, roughly a tenth of what it had paid for Claude Code and Codex. These are the company's own figures, presented in a conference talk.
Did Revolut train its own model?
Nothing in the published talk or the coverage describes fine-tuning on Revolut's own data. Revolut serves an open-weight model as released and adds its own context at run time. Training a model on company data is the next, separate step.
What does going headless mean for AI?
It means separating the tools people use from the model behind them. Revolut's coding agents call OpenAI-compatible endpoints, so the model behind them can be swapped, self-hosted or routed per task without changing how engineers work.
How many GPUs does it take to self-host a frontier-class open model?
Revolut serves about 2,000 engineers from 56 Nvidia B200 GPUs, the equivalent of seven eight-GPU systems. GLM-5.1's weights alone need roughly 380 GB of GPU memory, plus 6 to 26 GB per user for context, depending on precision.
Can a mid-sized company do what Revolut did?
The serving pattern scales down. A company can rent dedicated GPUs by the hour from inference providers, or start with per-token open-model endpoints and move to dedicated hardware once usage is steady. What does not scale down is the need to measure usage, latency and cost before switching.
Sources
- British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted. The Register, 3 Nov 2026. Accessed 9 Nov 2026.
- How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models: talk notes. Sparse Notes. Accessed 9 Nov 2026.
- Pricing: serverless inference, dedicated inference and GPU clusters. Together AI. Accessed 9 Nov 2026.
- AI Architects. OpenLazarus. Accessed 9 Nov 2026.
- Pricing. OpenLazarus. Accessed 9 Nov 2026.
Updated 9 Nov 2026: First published, based on Revolut's AI Engineer talk and The Register's report of 3 November 2026.
Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski