Revolut switches to Chinese GLM-5.1 model, cuts AI costs tenfold

From Gemini and Claude Code to GLM

According to Abitbol, Gemini's arrival in early 2025 trimmed coding time by 20% and delivery time by 10%. Revolut then turned to Anthropic's Claude Code, but with strict spending caps, after seeing engineers burn through thousands of dollars a month with little return. Claude Code, he said, did not have much impact at the time. Productivity figures jumped in April 2026 when the company moved to GLM-5.1, which has no token limits.

Paying for machines, not usage

Because GLM's weights are public, Revolut hosts the model itself on 56 Nvidia B200 chips. Billing no longer depends on consumption, so engineers can use the tool freely, in Abitbol's words, to "use AI without guilt". It is not free: Revolut pays for the hardware and must add more if demand outgrows it. He cited an average cost of 30 cents per million tokens.

Tuning took work. Time to first response fell from more than 10 seconds to 1.6 seconds on the same chips, while demand rose from 3 to more than 7 requests per second.

On concerns about Chinese models, Abitbol said the company downloads the weights and runs them on a US-based service rather than calling a Chinese one, keeping data in the EU or in countries with equivalent protection. The assistants work in an isolated environment with pre-approved operations and a human in control. He did not address risks from the model itself or the code it generates.

Since delivery times improved more than coding time, Revolut concludes that writing code was not the main bottleneck. Open models are now its priority, though Claude Code, Codex and Gemini remain available for some uses.