Revolut moves 2,000 engineers to self-hosted GLM at $0.30/M

Engineers at Revolut who used to ration Claude Code tokens now default to GLM, an open-weight model the bank hosts itself, AI operations lead David Abitbol said at AI Engineer NYC. The invoice, one-tenth of the old Anthropic and OpenAI spend, is Revolut’s own figure. It took roughly 56 B200 GPUs and tuning that he said started from “terrible” performance.

Blended cost comes to 30 cents per million tokens for 2,000 engineers. On the same GPU count, time to first token fell from over 10 seconds in April to around 1.6 by August. Codex and Claude Code stay available, kept as the exception.