Revolut Moves 2,000 Engineers to Self-Hosted GLM at Roughly a Tenth of Anthropic and OpenAI Costs
Revolut moved roughly 2,000 engineers onto a self-hosted deployment of the open-weights GLM model for coding and pushed Claude Code and Codex to an exception list, according to an X post by Guillermo Flor that cites David Abitbol, who leads AI operations at Revolut. The reported numbers are 56 Nvidia B200 GPUs, a blended cost of about $0.30 per million tokens — roughly one tenth of what Revolut had been paying Anthropic and OpenAI — and a 70% improvement in engineering lead time.
The detail that carries the most weight in the discussion is not the price itself but the workflow change behind it: the productivity gains appeared once engineers stopped rationing tokens.
What Revolut changed, and who reported it
Revolut swapped its default coding assistant for a self-hosted GLM deployment and demoted the frontier APIs it had been using to exception status. The claim comes from Guillermo Flor (@guilleflorvs) in a post dated November 2, 2026, attributing the operational specifics to David Abitbol, who leads AI operations at Revolut. The post states that roughly 2,000 engineers were moved, that Claude Code and Codex were placed on an exception list, and that the change covers coding work. The post also credits Oliver Molander for the tip.
No independent confirmation appears in the Hacker News thread, and every operational figure is secondhand — they originate from Abitbol's account as relayed by Flor. The closest thing to a primary source referenced in the discussion is a full recorded talk linked by Hacker News commenter swyx at https://www.youtube.com/watch?v=Ok6QdZHCcI8.
The cost and hardware numbers
Three figures anchor the story: 56 Nvidia B200 GPUs, a blended cost of roughly $0.30 per million tokens, and a 70% improvement in engineering lead time. Revolut's blended token cost is described as about one tenth of what the company previously paid Anthropic and OpenAI.
Hacker News commenters treated the headline cost as incomplete rather than decisive. gosen attempted a payback calculation: "So they've invested around $2.5M-3M into hardware + monthly operating costs (electricity, updates, cooling). I wonder how much did they spend on tokens to decide to make such a switch. I mean if they used to spend $500k-1M per month just on tokens then it would pay off in a year or so, assuming their infra will be stable and efficient." gosen also flagged an assumption baked into that math: the estimate only holds if the hardware is self-hosted in Revolut's own data center rather than rented from a cloud provider, which would be more expensive.
Two practical questions came up repeatedly:
- corford asked what "blended $0.30 per 1M tokens" actually means: "Is that input or output or average across the two?" The distinction matters for anyone benchmarking the figure against public GLM providers on OpenRouter.
- mtrovo asked how much operational work the cluster represents: "The relevant question is how easy is to setup 56x B200s in a cluster, are we talking about a adhoc infra team job or clicking a button on AWS?"
On the same X thread, NashBriggsy raised the staffing cost that rarely appears in token math: "30 cents per million is wild. the number i'd want next is how many people it takes to keep 56 b200s healthy, that's usually where the savings quietly go." koliber asked a supply-side question: "If I wanted to buy 56 B200's and had $3 million, when could I get delivery?"
Why removing token rationing mattered more than benchmark scores
The post's central claim is that the productivity result came from lifting a constraint rather than from a better model. Flor writes: "Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything."
Hacker News commenter littlecranky67 described the same dynamic from the opposite direction, observing a team where metered access produced caution instead of experimentation: "Most of them do not play with AI in their spare time, and they have been told to spend claude tokens wisely. Everybody is too afraid now to play with it and explore during work hours for fear of being token shamed." That commenter added that the same engineers still need the tooling skill set, because "most of their junior devwork can be replaced completely by AI."
This is the mechanism that generalizes beyond Revolut: when marginal token cost drops by an order of magnitude, the binding constraint shifts from budget to habit, and the payoff shows up in how broadly the tool gets applied rather than in per-task benchmark wins.
The quality question is unresolved
On the X thread, Kyle asked the obvious follow-up to a 90% cost cut: "Did the code get worse or just cheaper ?" 400_yen argued the gap is small for experienced users: "And to be honest the difference is negligible for seasoned engineers. Maybe if you're a junior Claude is hella smarter than GLM but if you know what you're doing GLM is enough for you. A few benchmark points less is nothing."
Hacker News commenter redmalang reported a similar production pattern with open-weight models, with a caveat about prompt specificity: "If you ignore the noise around 'cheap' etc, you might find that with a little extra effort to more precisely define what you need not only do these models generally work as well, and they will generally more closely follow instructions... They are less good on ambiguous prompts like 'fix all the problems' (which to my chagrin is a very common prompt it turns out..)"
aurreturn predicted the gap would reopen rather than close: "It's probably true that open weight models are 'good enough' for most software engineering tasks. But the goal posts will move and I think frontier labs will find new areas that have a need for absolute frontier."
hmokiguess framed the tradeoff as throughput over benchmark points: "I would sacrifice a few benchmark points for higher tok/s actually, especially if more bandwidth as well so I can scale up swarms, then I can likely compensate for any drop and get more mileage out of it."
Data sovereignty and regulation are separate drivers from price
Several commenters argued the cost saving understates why regulated companies self-host. jacquesm reported doing the same on much less powerful hardware: "We did the same thing, on much less powerful hardware but it has served us very well and we don't need to worry at all about the question whether or not the suppliers of inference services can be trusted long term with our data. The answer to that question in my opinion is 'absolutely not'."
torginus made the compliance case explicit: "there's little proof that the quality gap between open models and OAI/Anthropic stuff translates to business results, but you surrender all control of your data to a third party. The last cause might be a straight up blocker - certain financial regulations straight up prevent you from giving third parties to your data (esp outside national/EU jurisdiction), no matter what documents you sign." Revolut is a financial institution, which makes that constraint relevant independently of the token price.
Which organisations this actually applies to
The decision is most relevant to companies large enough to hit enterprise token pricing. injidup noted: "This is probably only relevant if you go over the 150 engineers putting you in the enterprise bucket and forcing you to pay by the token rather than the subsidized Claude code plans."
Other commenters reported similar moves at a different scale. world2vec described running "self-hosted GLM-5.3 via our flavour of pi.dev" while leaving teams discretionary budget for Claude Code, and observed that "most stuff doesn't require the latest and greatest LLM and can be done with 'lesser' LLMS by simply throwing a lot of tokens and compute to it." epolanski reported a mid-size German client with 100M in revenue "also moving lots of work to on premise LLMs." andy99 called the pattern "a sane version of what 'sovereign' AI can look like for many," adding that even if GLM is slightly behind, "this is likely close to the Pareto frontier." andy99 also suggested reselling spare capacity through a marketplace such as OpenRouter to improve hardware utilisation.
What the thread thinks it means for frontier labs
Flor's post frames the shift as "a pricing problem for every lab heading to IPO," on the logic that frontier models become the exception while open weights become the default. InsidiousWombat put a sharper edge on it: "If investors knew how powerful open models are the AI companies valuation would tank overnight." chr15m summarised the same sentiment as "Anthropic's missing last quarter results in a nutshell," and epolanski asked how model vendors sustain a moat at all.
The counterargument inside the thread is that self-hosting is a niche for large, regulation-bound buyers, and that open-weight parity holds for today's tasks rather than tomorrow's.
Caveats and open questions
The evidence base here is thin and should be treated as such. Everything comes from a single X post relaying a single operator's account; there is no published cost breakdown, no evaluation of code quality before and after the switch, and no confirmation that the 70% lead-time figure is measured rather than estimated. The model version is also unclear — one Hacker News commenter refers to "GLM-5.3," while the original post says only "GLM." The three numbers most worth watching for confirmation are the monthly operating cost of the 56-GPU cluster, the headcount required to keep it running, and whether the 70% lead-time improvement holds once the novelty of unmetered access wears off.