Revolut has moved roughly 2,000 engineers off metered frontier APIs and onto self-hosted GLM for everyday coding work. David Abitbol, who leads AI operations at the bank, said the setup runs on 56 Nvidia B200 GPUs inside Revolut's own infrastructure at a blended cost of about 30 cents per million tokens. That is roughly one tenth of what the company previously paid Anthropic and OpenAI. Claude Code and Codex have been pushed to an exception list, kept for the work where a frontier model still clearly wins.
The performance numbers came with the cost cut. Since the switch began in April, Revolut reports coding time down about 60 percent and engineering lead time down 70 percent. On the same GPUs, time to first token fell from more than 10 seconds to about 1.6 seconds between April and August, while requests per second climbed from around 3 to more than 7.
The most transferable lesson is operational, not technical. Abitbol says the biggest gains appeared only once tokens stopped being rationed. Engineers had been saving AI help for hard tickets and pointing it at little else. When the budget pressure vanished, they aimed agents at everything. He also advises teams to set a latency budget first, cap context at 256K tokens and keep cache hit rates above 90 percent for coding. For founders weighing AI spend, the story is blunt: open weights handled the default case and frontier APIs became the exception.