Which tenth?

The headline number deserves two readings because David Abitbol presented it in two different ways. At AI Engineer NYC on October 13, he said Revolut's invoice had fallen to one tenth of what it previously spent with Anthropic and OpenAI. At a separate October talk at NVIDIA GTC Berlin, presented by Together AI, he described the per-token cost of the equivalent frontier model as roughly ten times higher. Those are different comparisons. A tenth of the unit price, at much higher usage, does not necessarily produce a tenth of the total bill. No baseline figure, in dollars or tokens, accompanies either version, and there is no Revolut press release or engineering post to check them against. What exists is one executive's conference account, reported by The Register on November 3 and spreading since a summary post took off on X on November 2.

The $0.30 figure needs its version label kept on. The Register attaches it to GLM-5.1, the model Revolut self-hosted first: at $0.30 per million input and output tokens, it cost about a tenth of Codex or Claude Code. Coverage that attaches the same price to GLM-5.3, the version now in production, is merging two steps of the story. Talk summaries add that cache hits are free in the blended figure, another reason not to treat $0.30 as a directly comparable price for every input and output token.

Neither comparison establishes the fully loaded cost of running the deployment. The per-token figure excludes the cost of GPU capacity itself, the engineers who tuned the serving stack, and the monitoring and infrastructure around it. The invoice comparison also does not disclose enough detail to establish whether all those costs are included. On roughly the same 56 GPUs, Revolut reports load rising from about 3 to more than 7 requests per second between April and August, while time to first token fell from 14.26 seconds to 1.68 seconds. The company attributes the latency improvement to quantization, serving architecture and context-length tuning rather than additional hardware. Self-hosting at this level is an infrastructure program with a different cost structure, not a free model with no bill.

The gain that is not about the model

The productivity figures are the part most likely to be misread. In Revolut's own chart, the move to self-hosted GLM is where coding time improved by about 60 percent and lead time, from picking up a task to merging it, by about 70 percent. But two things changed in that step. The model changed, and the rationing ended. Abitbol's own explanation gives the larger role to the second: once tokens were effectively unlimited, engineers stopped reserving the tool for hard coding problems and let it work through debugging, Jira tickets and Confluence pages, which is where he says the real bottleneck sat. The productivity gains cannot be assigned to the model alone. Self-hosting changed the economics enough for Revolut to remove the token budgets, allowing engineers to use AI on work they had previously avoided. The company has not published a comparison separating the effect of better model access from the effect of ending rationing.

There is a market-level footnote that makes the case worth more than one bank's anecdote. Menlo Ventures' State of Enterprise AI report, published November 3, estimates open-source models at 19 percent of enterprise usage but only 4 percent of spend, and counts GLM-5.3 among the near-frontier models holding under 5 percent of API spend between them. Revolut's deployment illustrates one reason the gap can exist: self-hosted model usage generates little or no third-party API spending, even when engineers are using AI extensively. Menlo's figures do not establish how much of the gap comes from deployments like this. Menlo's figures are a venture firm's estimates, built on a survey of nearly 400 US enterprise decision-makers, and Menlo invests in several companies it covers. Read as a direction rather than a measurement, though, it matches what Revolut describes: open models are being used more than the money trail suggests.

Two parts of the setup will travel less well than the headline. The first is scale. Serving GLM at roughly 744 billion total parameters by Z.ai's reported figure, with about 40 billion active per token, to 2,000 engineers at interactive speeds is a commitment most companies cannot copy, and Revolut has not said whether its 56 B200s are bought, rented or run as dedicated capacity. The second is the part that did not change. GLM reads text only, so image work goes through a router to vision models, and Claude Code and Codex remain on call for the hardest problems. On the China question, Abitbol's position is that the risk sits in hosted APIs, not in weights: Abitbol's account is that Revolut uses publicly released weights on its own or regional infrastructure, applies human review and allow-lists, and does not send code to Z.ai's hosted service in China. Revolut has not published a separate compliance document detailing these controls. That is a defensible architecture. It is also a compliance workload, and it is part of the real price of the tenth.

Sources

  1. [1] The Register, "British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted", November 3, 2026 (read first-hand at Gate 1; the $0.30 GLM-5.1 figure, the $1,000–$10,000 monthly spends, the 56 B200 serving setup, the 256k context position, latency targets, router and exception arrangement)Read source
  2. [2] Cybernews, "Revolut Chinese AI model cuts coding costs by 90%", November 4, 2026 (read first-hand at Gate 1; today's account of the switch, human-review and US-infrastructure points, Ecosia parallel; note it attaches the $0.30 figure to GLM-5.3, which this article does not)Read source
  3. [3] CellCog, "Revolut Self-Hosts GLM: 2,000 Engineers, a Tenth of the Bill" (transcript-based write-up of the AI Engineer NYC talk, October 13, 2026, read first-hand at Gate 1; talk date, November 2 X spread, Gemini-step lead-time figure, chart attribution of the 60%/70% gains to the April GLM-5.1 step)Read source
  4. [4] The Independent AI, "Revolut moved ~2,000 engineers to self-hosted GLM at a tenth of the cost?" (claim check read first-hand at Gate 1; the two versions of "one tenth" across the New York and Berlin talks, no baseline published, unaudited status, Together AI's presentation of the Berlin session)Read source
  5. [5] Menlo Ventures, "2026: The State of Enterprise AI", November 3, 2026 (read first-hand at Gate 1; open-source 19% of usage vs 4% of spend, GLM-5.3 named among near-frontier models under 5% of API spend combined, methodology and investor disclosure)Read source
  6. [6] SemiAnalysis InferenceX, GLM-5 / GLM-5.1 model summary (read first-hand at Gate 1; GLM-5.1 architecture from the Z.ai model card: 744B vendor parameter figure, about 754B by checkpoint count, 40B active, MIT license)Read source