The PlaybookAI Economics

Revolut Dropped Claude Code and Codex to cut its token bill 90%.

Four accounts say Revolut moved 2,000 engineers to self-hosted GLM at a tenth of the token cost. Not one of them cites a primary source.

John TanNovember 2, 20265 min read

Over 48 hours at the start of November, one story about Revolut travelled across four accounts and arrived looking like established fact: the bank moved roughly 2,000 engineers off Claude Code and Codex onto a self-hosted GLM model, cut its cost per million tokens to about 30 cents, and now pays something like a tenth of what it paid the frontier labs.

It may all be true. The problem is that nobody carrying it cites a source, and the account that corroborates the numbers is the one that arrived last.

4

Accounts carrying the story in 48 hours, 1 to 2 November.

0

Primary sources cited by any of them. No statement, no recording, no link.

74%

Gap between two valuations relayed eight days apart. $115B, then $200B.

What all four actually agree on

Strip out every figure and something real is left. An engineering organisation of serious size has moved its coding agents off OpenAI and Anthropic onto open weights it hosts itself, on cost grounds rather than capability grounds. That is worth your attention, because it is the first such case at a company that does not sell AI.

The earlier ones did sell AI. Cursor shipped Composer 2 and Harvey shipped Tenet, both in late September, both companies with the inference volume and the ML staff to justify it. The argument at the time was that self-hosting needs volume and model capability, which is why it showed up at AI companies and not elsewhere. A bank is neither of those things. If a bank has done it, the threshold moved.

Notice that none of that requires a number. The direction is the finding. The arithmetic is where it falls apart.

Oliver Molander
Oliver
Molander

"Revolut reportedly only started seeing meaningful AI productivity gains after moving away from expensive OpenAI and Anthropic models to self-hosted GLM. Decagon reportedly runs ~90% of its workloads on open weight models."

Oliver Molander  ·  X  ·  November 1, 2026

That is the first account into the story, and the only one that hedges. Two "reportedly"s, no figures at all, and a second company named alongside so the pattern carries rather than the anecdote. It is also the version nobody reshared, because it has nothing quotable in it.

Figure 1. Who carries which number
The figures appear downstream, not upstream.
ACCOUNT DATE FIGURES IT STATES SOURCE CITED Molander 1 Nov none none Account 2 2 Nov $200B · 2,000 engineers · $1k–$10k/mo none Account 3 2 Nov 30c / Mtok · 1/10th the cost "Revolut CTO" Account 4 2 Nov 2,000 · 56 B200s · 30c · 1/10th a named person The only account that hedges is the one with no numbers in it. Every figure enters the story at the second telling or later. And the only account that repeats earlier figures is the last one to arrive, which adds one nobody else has.
Four X posts dated 1 and 2 November 2026, read 3 November 2026. Figure by nativefirst.

Where the numbers come from

The 30 cents per million tokens and the "tenth of the cost" first appear in an account quoting Molander's post. Neither figure is in the text being quoted. They are the quoting account's own addition, attributed to a "Revolut CTO" who is not named and a venue that is not given.

The $200B valuation, the 2,000 engineers and the $1,000 to $10,000 a month per engineer come from a different account, which carries none of the cost figures. So the two most-quoted posts in this story share exactly nothing.

Guillermo Flor
Guillermo
Flor

"BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list. Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic"

Guillermo Flor  ·  X  ·  November 2, 2026

This is the version that reads like reporting. It has a hardware count, a named person with a job title, and the word BREAKING. It is also the fourth telling, and the 56 B200s appear nowhere else.

The name that makes it look sourced

David Abitbol does appear earlier in this story, and that is the detail worth being careful about. One of the accounts flags a conference talk in which Abitbol discusses productivity figures of 60 to 70 percent using data from 2,000 engineers. Nothing in that post says he was speaking about Revolut, and his role is not stated.

Two posts mention 2,000 engineers. One mentions Abitbol. By the fourth telling he has become the head of AI operations at Revolut and the source of the cost figures. Nobody invented that on purpose. It is what happens when two fragments share a number and a relay needs an attribution.

A named source with a title is the single most persuasive thing you can add to a claim, which is exactly why it is the thing to check hardest.

The valuation nobody can pin

Eight days before all this, a podcast relay put Revolut at $115 billion. The 2 November post puts it at $200 billion. Both are unsourced, 74% apart, and nothing reconciles them. If the basic valuation of the company cannot survive a week, the per-token cost of its coding agents is not the number to build a strategy on.

How to read a claim like this

Find the first telling, then check what it actually says. Stories gain figures as they travel. Walk back to the earliest post and read it. Here, the earliest version has no numbers and says "reportedly" twice, which tells you everything about what the later confidence is made of.

Treat repetition as echo until you can prove it is corroboration. Two accounts stating 30 cents looks like confirmation. It is one claim and one reshare. Corroboration means two parties who could each have measured it independently, and that is rare enough that you will usually find there is only one. The related trap is one number quoted in two directions, which is exactly what happened to Sweden's AI adoption rate.

Separate the direction from the arithmetic. You can act on "a bank moved its coding agents to self-hosted open weights" today. You cannot act on "30 cents per million tokens" until somebody says where it came from. The first is the strategically interesting part anyway.

Run the number on your own fleet before you quote anyone's. Self-hosting trades a per-token bill for fixed capacity, staff and utilisation risk, so the break-even is specific to your volume. The same discipline applies to the headline token numbers everybody is quoting this month, and GLM's own published results are checkable in a way this story is not.

Every company will not end up copying Revolut, whatever the posts say. Some will, once somebody publishes a number with a method attached.

Ask who measured it.

Portrait of John Tan
John Tan ↗

Founder of nativefirst.ai. Runs everything on agents and writes down what works.

Sources
1Four X posts dated 1 and 2 November 2026, read 3 November 2026. The earliest, by Oliver Molander on 1 November, states that "Revolut reportedly only started seeing meaningful AI productivity gains after moving away from expensive OpenAI and Anthropic models to self-hosted GLM" and that "Decagon reportedly runs ~90% of its workloads on open weight models." It carries no figures and names no source.
2The $200 billion valuation, the 2,000 engineers and the $1,000 to $10,000 per engineer per month come from a second account on 2 November, which carries none of the cost figures. The 30 cents per million tokens and the "1/10th the cost" come from a third account on 2 November, quoting Molander's post; neither figure appears in the text it quotes, and both are attributed to an unnamed "Revolut CTO" at an unnamed venue.
3The fourth account, Guillermo Flor on 2 November: "BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list. Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic." The 56 B200s appear in no other account, and Abitbol's role at Revolut is not established by any of the four posts.
4David Abitbol appears earlier only in the third account's separate reference to a conference talk discussing productivity figures of 60 to 70 percent using data from 2,000 engineers. Nothing in that reference states he was speaking about Revolut.
5The $115 billion valuation is from David Senra's podcast episode with Nik Storonsky, released around 25 October 2026, and is Senra's own figure rather than a sourced valuation. Neither it nor the $200 billion figure is traceable to Revolut or to a named publication.
6The earlier self-hosting cases referenced are Cursor's Composer 2 and Harvey's Tenet, both 23 September 2026. No Revolut statement, recording, filing or named publication was found for any claim in this story.

Read next

AI is moving really fast. Keep up.

Read the Playbook