What did Revolut replace Claude Code with?

About 2,000 Revolut engineers adopted an open-weights coding model from Z.ai in place of Claude Code. The Z.ai model ran on 56 Nvidia B200 GPUs at about $0.30 per million tokens.

How did token limits affect AI coding use at Revolut?

Strict budget limits on Claude Code led Revolut engineers to reserve tokens for only the hardest problems. After switching to the Z.ai model without the token cap, Revolut engineers began using the model for debugging.

What productivity improvement did Revolut report after switching AI coding models?

Revolut’s internal research reported an approximately 60% reduction in lead time following the switch from budget-limited Claude Code to a Z.ai coding model. The post defines lead time as the interval from picking up a feature ticket to having the feature merged and ready to ship.

What should engineering teams learn from Revolut’s AI coding switch?

The post’s lesson from Revolut’s AI coding switch is that token budgets can constrain how engineers use coding models, beyond the direct cost of tokens. Removing pressure to ration tokens can broaden AI use for routine coding work and leave more attention for features and architecture improvements.

When Engineers Stopped Rationing Tokens: The Real Lesson from Revolut’s AI Switch

Revolut's AI story is being told as "open model beats Claude at a tenth of the price." What’s more interesting is what happened when their engineers were freed up from rationing tokens.

About 2,000 Revolut engineers started using an open-weights model from AI lab Z.ai as their go-to coding model, running on 56 Nvidia B200 GPUs at about $0.30 per million tokens. It replaced Claude Code which came with strict budget limits, so engineers were saving their tokens for only the hardest problems.

Without the cap, they started using the new model or debugging.

Revolut's internal research shows their time spend coding was cut by about 60% ('lead time' being time from picking up a feature ticket to having it merged and ready to ship).

Teams that aren’t worried about hitting their monthly token allotment can focus on more features and architecture improvements that scale.

Frequently asked questions

What did Revolut replace Claude Code with?: About 2,000 Revolut engineers adopted an open-weights coding model from Z.ai in place of Claude Code. The Z.ai model ran on 56 Nvidia B200 GPUs at about $0.30 per million tokens.
How did token limits affect AI coding use at Revolut?: Strict budget limits on Claude Code led Revolut engineers to reserve tokens for only the hardest problems. After switching to the Z.ai model without the token cap, Revolut engineers began using the model for debugging.
What productivity improvement did Revolut report after switching AI coding models?: Revolut’s internal research reported an approximately 60% reduction in lead time following the switch from budget-limited Claude Code to a Z.ai coding model. The post defines lead time as the interval from picking up a feature ticket to having the feature merged and ready to ship.
What should engineering teams learn from Revolut’s AI coding switch?: The post’s lesson from Revolut’s AI coding switch is that token budgets can constrain how engineers use coding models, beyond the direct cost of tokens. Removing pressure to ration tokens can broaden AI use for routine coding work and leave more attention for features and architecture improvements.