Revolut moved about 2,000 engineers to GLM, deployed in-house, leaving Claude Code and Codex as exceptions. David Abitbol, ​​who is responsible for the operation of artificial intelligence in the company, spoke about this in the report. The model runs on 56 NVIDIA B200s, a million tokens cost an average of $0.30, and the total bill is about 10 times less than what Revolut paid Anthropic and OpenAI. According to internal measurements, the time from entering a task into production to merging ready for deployment was reduced by 70%. According to Abitbol, ​​the impact appeared when tokens stopped being saved and engineers started delegating some tasks to AI. Gartner analysts reach a similar conclusion in a new report, the Register writes: 9 out of 10 steps in enterprise agent processes are routine, and open models deal with them on par with closed ones. If you take into account the text and repetition, the savings compared to the front-end model is 5-6 times per task.

Revolut moved about 2,000 engineers to GLM, deployed in-house, leaving Claude Code and Codex as exceptions. David Abitbol, ​​who is responsible for the operation of artificial intelligence in the company, spoke about this in the report. The model runs on 56 NVIDIA B200s, a million tokens cost an average of $0.30, and the total bill is about 10 times less than what Revolut paid Anthropic and OpenAI. According to internal measurements, the time from entering a task into production to merging ready for deployment was reduced by 70%. According to Abitbol, ​​the impact appeared when tokens stopped being saved and engineers started delegating some tasks to AI. Gartner analysts reach a similar conclusion in a new report, the Register writes: 9 out of 10 steps in enterprise agent processes are routine, and open models deal with them on par with closed ones. If you take into account the text and repetition, the savings compared to the front-end model is 5-6 times per task.