Revolut runs 2,000 engineers on self-hosted GLM at a tenth of its Claude bill

That matters beyond one fintech's procurement spreadsheet. A regulated lender with more than 80 million customers in 45 markets now runs its default coding model on its own hardware, which hands every engineering team a concrete price comparison against the frontier labs. It also gives regulators a live case study in running Chinese-origin weights on Western infrastructure. What Revolut has not said is what the chips cost to buy or rent, or how much engineering time the tuning took.

Fifty-six GPUs for 2,000 engineers

The setup is small. Revolut runs GLM, a model from the Chinese lab Z.ai with about 800 billion parameters, on roughly 56 Nvidia B200 chips. Abitbol put the cost at about 30 cents per million tokens, with a flat bill replacing unpredictable month-end invoices.

Getting there took work. Out of the box, inference was "terrible," Abitbol said. After optimization, time to first token fell from more than 10 seconds to 1.6 seconds even as demand rose 76 percent. He said that above five seconds, "my Slack inbox starts to blow up". The context window is capped at 256,000 tokens. "Long context is a scam," he said. GLM handles text only, so a router sends image tasks to Gemini, and Claude Code and Codex remain available as the exception rather than the default.

Coding time fell 60 percent and lead time 70 percent across a sample of about 2,000 engineers. Abitbol drew a lesson from the gap between those two numbers. "The bottleneck for engineering is not necessarily coding," he said, noting that engineers spend much of their time searching inboxes, Jira and Confluence.

Gartner puts the saving at a twentieth

Revolut is not alone in doing the arithmetic. The Register reported on 3 November that Gartner reckons open models can handle 90 percent of agentic work at a twentieth of the cost of a frontier LLM. That is an analyst estimate rather than a customer bill, but it points the same way as Revolut's numbers.

The price gap in numbers

Closed models still lead on raw capability. Gemini 4 Carbon scores 59 on Artificial Analysis Intelligence Index V5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. It also costs $7.76 per task on that test.

Put that next to the open-weight end of the chart. DeepSeek V4.1 Flash 1017 scores 32 on the same index and costs about 13 times less per task, according to Tom's Hardware's reading of the Artificial Analysis figures. On Arena.ai's Pareto frontier chart, GLM 5.3-Flash and DeepSeek V4 Pro take over at lower price points.

The frontier labs still win the top of the ranking. They no longer win the middle.

For Revolut's engineers, the change is less about price per token than about not counting them. Token limits are gone, so nobody rations usage near the end of a quarter or on a demanding task. That is a workflow change as much as a finance one.

What open weights do not fix

The model's origin remains sensitive. "There's been a lot of stigma in using Chinese open source model," Abitbol acknowledged. Revolut uses only the openly available weights and runs them on infrastructure in the United States rather than through the developer's inference endpoint in China. Data is meant to be processed only in jurisdictions with equivalent protection, such as the US, the EU, Britain and Israel, and providers must have zero data retention by default.

The agents do not act alone. They run in a sandbox with approval rules that decide which commands are allowed. "They're always with a human in the loop," Abitbol said.

Mistral makes the sovereignty case explicitly, pitching Mistral Large 4 as a Western-developed model that companies can run inside their own firewalls. Nvidia, for its part, sells the GPUs either way. Jensen Huang is happy to take your money whether you use open or closed models, The Register reported on 3 November.