Revolut says making self-hosted open models its coding default reduced its inference bill to roughly one-tenth of earlier Anthropic and OpenAI spending. David Abitbol, who leads its AI operations, described the deployment at AI Engineer NYC on October 13.
He reported a blended cost near $0.30 per million tokens, serving roughly 2,000 engineers with GLM-5.3 on 56 B200 GPUs. Claude Code and Codex remain available for exceptions.
The model itself comes from Z.ai. Its published GLM-5.3 model card provides weights and documents deployment through frameworks including vLLM and SGLang. Running those weights on selected infrastructure is different from announcing a new foundation model developed by Revolut.
What the savings figure establishes
The figures describe company-reported deployment economics. The presentation does not provide a separately itemized total cost of ownership or an independently audited savings study.
For another engineering organization, the comparison needs a defined workload and accounting boundary. A token rate, a monthly inference invoice and the full cost of operating a platform measure different things. Hardware or hosting commitments, utilization, staffing and support can change the result. A lower unit cost also permits more usage, so cost per completed task deserves attention alongside spending.
Control the system around the model
The UK NCSC’s guidance on agentic AI says model safeguards should not be treated as complete protection. It recommends matching autonomy to consequences, enforcing additional controls, protecting activity logs and retaining the ability to halt the wider system.
For a coding deployment, approval boundaries and access to repositories and credentials remain separate design decisions from where inference runs. Evaluating the serving bill does not answer whether the agent can perform an unauthorized action.
