REVOLUT TAKES AI OFF RATIONS

The real news

The story, straight

An AI coding assistant reads instructions and produces text, including computer code. Services often charge by tokens: small chunks of text processed or generated.

Put engineers on a token budget and they save their allowance for difficult tasks or quarter-end. Congratulations: your productivity tool now needs a special occasion.

Self-hosting means downloading a model's weights, the files containing its learned settings, and running them on hardware you operate. You pay for machines and operations instead of buying each text chunk from a model provider.

At October's AI Engineer conference in New York, Revolut's David Abitbol described making self-hosted GLM models from Z.ai the default. The Register reported the presentation on 3 November 2026.

Revolut says its AI coding inference bill, the cost of running model requests, fell roughly 90% in under a year.

That compares bills, not identical workloads. It excludes hardware, cloud, staffing and monitoring costs.

The computers have not joined a volunteer programme.

The interesting change was behaviour. Abitbol said cheap, unlimited tokens removed the need to ration requests.

The biggest gains came from debugging and handling Jira and Confluence tickets: finding software faults and dealing with work requests and documentation. The paperwork finally got something to chew on besides humans.

Claude Code and Codex remain available for exceptional work. GLM is text-only; image requests go elsewhere.

This is a change of default, not an eviction notice.

The lesson: improve the work that slows people down, not just typing speed. Running your own model can help if you can keep the machines tuned and busy.

Owning the buffet does not make electricity free.

Source: https://www.fstech.co.uk/fst/Revolut_cut_AI_coding_costs_10x_thorough_self_hosted_models.php