Revolut reports 90% AI coding cost cut after self-hosting open-weight GLM

Revolut cut its AI coding inference costs by roughly 90% after moving its roughly 2,000 engineers off frontier-model APIs as the default option and onto a self-hosted, open-weight model. The figures come from a dev.to write-up of a talk by David Abitbol, who leads AI operations and infrastructure at the British fintech, delivered at the AI Engineer conference in New York; the talk was posted on 29 October 2026 and the write-up appeared on 3 November. Coding time reportedly improved by about 60% over the same period.

The path to self-hosting

According to the timeline Abitbol presented, the rollout went through several distinct phases:

• Early 2025: Gemini was deployed company-wide and cut coding time by around 20%.
• Early 2026: Claude Code arrived under strict per-engineer budgets, because some engineers had been running up token bills of $1,000 to $10,000 a month with little measurable return.
• April 2026: Revolut began serving open-weight models on its own hardware, starting with GLM-5.1 from the Chinese lab Z.ai, and removed token limits entirely.
• Later: the company upgraded to GLM-5.3, a roughly 754-billion-parameter model with about 40 billion parameters active per token, served from around 56 Nvidia B200 GPUs.

Claude Code and OpenAI's Codex have not been eliminated. They now sit on an exception list and are reserved for work the self-hosted setup cannot handle; GLM handles text only, so UI-related tasks and the hardest problems still route to frontier models.

The numbers Revolut reported

The dev.to write-up stresses that these are Revolut's own figures and have not been independently verified:

• Blended serving cost of about $0.30 per million tokens, roughly a tenth of the previous Anthropic and OpenAI invoices.
• Around 60% improvement in coding time and 70% in lead time from pickup to merge, across roughly 2,000 engineers.
• 76% more requests per second with no additional GPUs after inference tuning, and time-to-first-token down 88%, from 14.3 seconds to 1.7 seconds.
• The GPU fleet held steady at 56 B200s from April through August even as usage grew.

Unlimited tokens over smarter models

The write-up's central reading of the data is that the speed gain came less from model quality than from removing the meter. Engineers previously rationed their AI usage because every API call carried a cost; once tokens were effectively free at the margin, behaviour changed, and that behavioural shift drove the improvement. On this account, GLM is not a stronger model than the frontier options it replaced for daily work — it is one engineers could use without weighing each request against a bill.

There is a compliance dimension as well. Revolut runs the model weights on US-based infrastructure rather than calling Z.ai's hosted endpoint in China, which keeps engineering data on hardware the regulated bank controls.

The caveats

The write-up is candid about the limits of the case study:

• The switch to a new model is not isolated from the removal of token limits, so the headline figures bundle several variables together.
• The $0.30-per-million-token figure excludes cloud spend, staffing and observability costs.
• A 56-GPU B200 fleet is substantial infrastructure, and a small team cannot replicate it directly.
• The throughput gains came from sustained inference-tuning work, not simply from downloading weights.

Why it matters

This is one of the few documented, large-scale enterprise case studies of AI cost control through self-hosting, and it lands at a moment when many engineering organisations are watching AI line items climb. It also challenges the assumption that model quality is the dominant variable in coding-assistant returns: Revolut's reported gains came from making tokens cheap enough that engineers stopped rationing them.

The playbook the write-up draws out — open-weight models for routine work, frontier models for exceptions — scales down even where a 754-billion-parameter model does not. Smaller openly licensed models and inexpensive hosted inference for open weights put a version of this strategy within reach of much smaller teams, though Revolut's specific numbers, being self-reported and tied to its own infrastructure, should be read as one company's experience rather than a guaranteed template.