Revolut Made a Self-Hosted GLM Model the Default Coding Tool for 2,000 Engineers
Neobank Revolut serves GLM-5.3 on 56 Nvidia B200 GPUs to roughly 2,000 engineers and claims to pay ten times less than with Anthropic and OpenAI. Claude Code and Codex remain in use, but only as exceptions.
Neobank Revolut has turned an internally hosted open model into the default coding tool for its developers. David Abitbol, head of AI operations, explained the move at the AI Engineer conference in New York in October, and The Register covered his presentation on November 3, 2026. Revolut serves GLM-5.3, a model from Chinese lab Z.ai, to roughly 2,000 engineers on 56 Nvidia B200 GPUs. Its AI coding bill is reportedly ten times lower than with Anthropic and OpenAI.
From strict budgets to unmetered tokens
David Abitbol explained that Revolut initially granted access to Gemini in early 2025, followed by Claude Code at scale, operating under strict budget caps: some engineers were spending “$1,000 a month, $10,000 a month” with no clear return. The company therefore began self-hosting open models around April 2026, starting with GLM-5.1 and later moving to GLM-5.3. The average cost dropped to around 30 cents per million tokens, shifting billing to a predictable fixed cost with no more surprises at the end of the month.
The biggest payoff, he noted, does not even stem from writing code. With unlimited tokens, engineers also rely on the model for debugging, Jira tickets, and Confluence documentation. Revolut reports 60% less time spent coding and 70% shorter delivery times. These remain internal figures, and they coincided with lifting token rationing, making it difficult to isolate the model’s intrinsic impact. Claude Code and Codex have not disappeared entirely: “frontier becomes the exception,” reserved strictly for the most complex tasks.
The settings that make a difference
The talk served primarily as a technical post-mortem:
| Setting | Revolut’s choice |
|---|---|
| Model | GLM-5.3, text-only |
| Hardware | roughly 56 Nvidia B200s for 2,000 engineers |
| Latency target | under 3 s to first token |
| Throughput | around 200 tokens/s (close to 800 in the evening) |
| Context | 256,000 tokens, not one million |
| Quantization | FP4 on Blackwell GPUs |
| Images | routed to a different model |
“Long context is a scam,” Abitbol summarized: even if the model supports one million tokens, 256,000 are plenty. According to him, between April and August, time to first token dropped from about 10 seconds down to 1.6 seconds with the exact same GPU footprint despite surging demand, achieved through optimization rather than adding hardware. The endpoints maintain compatibility with the OpenAI API and are plugged directly into OpenCode or Codex.
For a bank operating under European regulations, he underscored two crucial guardrails: a sandbox ensuring a human signs off on actions, and comprehensive audit logs in case of incidents. Regarding Chinese models, he argued that risks stem from APIs hosted in China, not from downloading weights and running them on servers in the United States or Europe.
Our take
Revolut’s case demonstrates how the economics of coding agents fundamentally shift at enterprise scale: beyond a few thousand heavy users, renting GPUs and serving an open model can cost significantly less than seat-based subscriptions. However, doing so requires a dedicated team capable of fine-tuning inference, and a cluster of 56 B200 GPUs remains out of reach for most businesses. The reported productivity gains should be taken with a grain of salt: they primarily reflect the impact of removing token caps. For Anthropic and OpenAI, the takeaway is bittersweet: while their models remain the gold standard for difficult tasks, daily bulk volume can easily migrate toward open alternatives.
Sources
- AI Engineer — transcript of “Shipping like a startup, regulated like a bank,” David Abitbol (AI Engineer NYC 2026) Primary source
- AI Engineer — Revolut talk page Primary source
- The Register — British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted (November 3, 2026) Secondary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
