Revolut cut AI costs 90% by replacing expensive OpenAI and Anthropic models with self-hosted Chinese GLM models for 2,000 engineers.
It reports 60% faster coding and 70% shorter engineering lead times (from starting work to being ready to merge).
The bigger insight was cheaper AI removed usage limits, letting engineers automate work beyond coding like email, Jira, documentation and meetings.
The productivity gains came from both writing code faster and speeding up everything around it. https://youtu.be/Ok6QdZHCcI8?is=TotDGGT4EiB32Suw
How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models - David Abitbol, Revolut
YouTube
This is why people who hold out hope that token prices staying high will prevent AI from disrupting their jobs are misguided.
A focused effort on optimizing token costs away from frontier models can bring token costs down by 10x. As these stories spread, this will happen more often.