Revolut has stopped using the Anthropic and OpenAI APIs as the default coding tool for roughly 2,000 engineers and replaced them with an open-weight model from the GLM family, built by Chinese lab Z.ai, served on 56 NVIDIA B200 graphics processing units (GPUs) reserved for that use. According to figures disclosed by David Abitbol, the fintech’s head of AI operations, the combined cost runs around $0.30 per million tokens, roughly a tenth of what it used to pay the two US labs.
Revolut’s self-hosted inference in 20 seconds
- About 2,000 engineers now code by default with a GLM model from Z.ai served on 56 B200 GPUs.
- The stated cost is $0.30 per million tokens, not including staff, cloud, or observability.
- Claude Code and Codex remain available only as an exception.
- The figures come from the company itself and don’t reflect total cost of ownership.
The figures come from a talk Abitbol gave, published by the AI Engineer channel on 10/29/2026, as reported by StartupHub.ai and AIChatDaily, and summarized by aggregator daily.dev drawing on several sources. This outlet hasn’t located a post of Abitbol’s own on LinkedIn or X with these numbers, so they should be read for what they are: company-provided figures, without external audit. It’s also unclear exactly which version of the model Revolut uses. daily.dev says GLM-5.3, other reports simply say “GLM models,” and Together AI has published a case study of an unnamed “global fintech” serving GLM-5.2 on 56 B200s split across 14 replicas of four GPUs each. The matching numbers are striking, but Together doesn’t identify its client.
The workload split looks like this: the open model is the default option, Claude Code and Codex remain available as an exception, and tasks involving images are routed to Gemini or to GPT-6 Luna, the economy model OpenAI launched in September. The self-hosted path is therefore text-only. Abitbol credits this phase, which began in April 2026, with a 60% improvement in coding time and a 70% improvement in engineering lead time, the time it takes a change to reach production. The chart he showed doesn’t separate the model’s effect from other simultaneous changes, and the talk itself acknowledges that the previous phase, with Claude Code and very tight token budgets, had little impact.
What a Million Tokens at $0.30 Actually Costs
The $0.30 figure is a combined input-and-output cost that, according to reports on the talk, leaves out cloud spending, the staff who maintain the system, and observability tooling. In other words, it isn’t the TCO (total cost of ownership), which is what matters to whoever signs off on the budget.
AIChatDaily has made its own estimate, not Revolut’s, using market prices of $6 to $8 per GPU-hour. Under that assumption, renting 56 B200s would cost between $245,000 and $327,000 a month, about $120 to $165 per engineer, and buying them outright would run around $2.7 million in hardware alone. On top of that comes the team that tunes and operates the platform, which is exactly what the $0.30 figure leaves out.
Checking the claimed “tenth of the cost” against public pricing isn’t straightforward. As of 11/05/2026, Anthropic charges $2 per million input tokens for Claude Sonnet 5.5, $0.10 for cached reads, and $10 per million output tokens; Claude Opus 5.5 costs $4, $0.20, and $20. OpenAI cut the price of GPT-6 Sol on 09/22/2026 to $2 for input, $0.20 with caching, and $10 for output. In agentic coding, input dominates: in Together AI’s case study, the median request ran around 81,000 input tokens. The real average price therefore depends on how much input is served from cache and on any negotiated discounts, and Revolut hasn’t published either figure. A tenth of the cost would imply a previous cost of around $3 per million, a plausible figure at those prices, but not one that can be verified from the outside.
Tuning Inference, the Part That Never Shows Up on the Bill
The most useful part of the talk for an infrastructure team is the tuning work. According to StartupHub.ai, Abitbol summed up the starting point with a line that translates roughly as “inference out of the box is terrible.” After tuning, time to first token dropped from more than 10 seconds to 1.6 seconds, and throughput went from 3 to more than 7 requests per second. The company says that’s an 88% performance improvement to handle 76% more demand without adding GPUs.
Other technical details: using NVFP4, the 4-bit format on Blackwell GPUs, reportedly cut the GPU footprint by around 40% compared with FP8 with no loss of accuracy, again according to the company. Context was capped at 256,000 tokens, and the model needs tensor parallelism across eight GPUs (TP8) to fit the key-value cache. The service-level targets set are under 3 seconds to first token, around 200 tokens per second in interactive use, and peaks of 800 tokens per second from 5 p.m. onward.
The setup carries risks. In the case of the unnamed fintech Together AI describes, there were episodes of capacity shortages with requests queued for one to three minutes. That’s the kind of problem a public API absorbs on its own and that, with dedicated capacity, becomes the customer’s responsibility. That’s why this kind of deployment fits better with a FinOps discipline applied to cloud cost than with a simple price-per-token comparison.
Stack Sovereignty and Open Questions for a Regulated Bank
Revolut isn’t a startup. In July 2026 it confirmed a secondary share sale that, according to press reports, valued it at $115 billion, 53% more than the $75 billion valuation from its November 2025 round. Bloomberg put the price at $2,017 a share, though Dealroom cites around $2,118. Because it was secondary, no new money went into the company. That scale explains why it can afford its own inference team, something a small company could hardly justify.
On security, StartupHub.ai reports that every agent run has human oversight and runs in an isolated environment with allow-listed commands, that inference stays within countries with data-protection equivalence (the United States, the EU, the UK, and Israel), and that the GPU providers don’t retain data by default. By running the weights on dedicated capacity rather than through Z.ai’s API, the code never passes through the Chinese lab’s servers.
Some questions remain without a public answer. GLM-5.3’s license isn’t MIT, unlike the Flash version: it’s a proprietary license that, according to IntuitionLabs, requires operators that exceed $10 billion in aggregate revenue and offer the model as a service to pass a Z.ai security review. There’s no indication of whether that clause applies to Revolut’s internal use. Nor have any statements turned up from British or European regulators on the use of models from Chinese labs at banking-licensed institutions, such as the one Revolut obtained in the UK this year.
For cloud providers and labs, the case opens a debate that for now rests on opinions. According to daily.dev, @guilleflorvs, who first broke the story, argues the improvement came once token rationing stopped, and that if open models become the norm, labs preparing to go public will have a pricing problem. @omarsar0 argues that companies should own their own intelligence stack. This outlet already reported on how GLM-5.2 started competing in coding against frontier models, but there are no independent measurements comparing the quality of Revolut’s code before and after the switch.
Frequently Asked Questions
What model does Revolut use for coding?
An open-weight model from the GLM family, built by Z.ai. daily.dev identifies it as GLM-5.3, while other reports don’t specify the version.
How much does self-hosted inference cost Revolut?
The company cites about $0.30 per million tokens, but that figure excludes cloud, staff, and observability. External estimates put renting 56 B200s at between $245,000 and $327,000 a month.
Has Revolut stopped using Claude Code and Codex?
Not entirely. They remain available as an exception, and tasks involving images are sent to Gemini or GPT-6 Luna.
Is this a useful strategy for small companies?
It requires a team capable of operating and tuning dozens of GPUs, so, according to daily.dev, startups are likely to stick with APIs.
Sources: StartupHub.ai, AIChatDaily, Together AI, Tech.eu, Sifted, Dealroom, Gigazine, IntuitionLabs, Claude API pricing, Implicator, and daily.dev (summary: “Revolut moves about 2,000 engineers to self-hosted GLM for coding”).
This article is also available in Spanish: “Revolut cambia la API por 56 GPU B200 y un modelo abierto para programar”, on revistacloud.com.
