Revolut cut its AI coding bill by about 90% by running an open model on its own chips. Before you renew a coding-agent contract, test an open model on your own tickets

Revolut, the British fintech, says it cut the cost of AI coding help by roughly 90% in under a year. It did it by moving most of that work off the commercial coding agents, Claude Code and OpenAI's Codex, and onto GLM, a family of open-weight models from the Chinese developer Z.ai. Open-weight means anyone can download the model and run it on their own hardware. David Abitbol, who leads operations and infrastructure at Revolut, described the switch at the AI Engineer conference in New York, as reported by The Register(opens in a new tab) on November 3.

The numbers are specific. Revolut serves(opens in a new tab) GLM-5.3 to about 2,000 engineers on 56 Nvidia B200 chips. Self-hosting GLM cost it about $0.30 per million tokens (a token is roughly three quarters of a word), about a tenth of what the commercial agents cost. Before the switch, Abitbol said, some engineers were spending $1,000 to $10,000 a month 'with not so much ROI.' Since rolling out AI coding tools, the company reports coding time improved by about 60% and lead time by about 70%. The bigger gain, he said, came from letting engineers use the model freely on debugging and ticket work instead of rationing tokens for the hardest jobs.

This is the capability shift worth noting. A year ago, open models were the cheap option you accepted lower quality for. A large, regulated fintech now runs one as its default coder. But read the limits. Revolut published no quality comparison against the commercial models, only speed and cost. GLM is text only, so screenshots and interface work still go to Gemini or OpenAI through a router. And 56 top-end chips is a fixed cost that pays back at 2,000 engineers, not at twenty. Abitbol also flagged the obvious question with a Chinese model: his answer was to run it on his own or regional hardware, not through an API hosted in China.

The other bank in today's news went the other way. Goldman Sachs has pulled advanced coding tools(opens in a new tab) from most traders and bankers and kept them for engineers, citing productivity and risk. The model is not the decision. Who gets it, for which work, at what price is.

If you run an engineering team that pays for coding-agent seats, take one week of real debugging and ticket work, run it through a hosted open model on infrastructure you trust, and compare cost and accepted fixes before your next renewal.