How Revolut cut its AI coding bill 90%

Some headlines said Revolut cut out Claude to save money. That isn't quite what happened, and the real playbook is more useful for any team running AI coding tools.

Revolut made an open-weight model, GLM, the default for its coding agents. It says the blended cost is about 30 cents per million tokens. It routes each job to the model that handles it best, so code stays on GLM while screenshots and design work go to Gemini or GPT. It also tuned its serving setup, which raised the requests its 56 Nvidia B200 chips could handle by 76 percent.

Revolut reports coding time improved about 60 percent and lead time about 70 percent after the switch. Those are its own numbers, and the timing alone doesn't prove the model caused all of it. A small team is better off renting compute than buying B200s.

What you can copy this week: set a monthly cap on every AI coding seat, pay top-model prices only for hard work, and track one metric like lead time before and after each change. What does your team's monthly AI coding spend look like right now?