Revolut says self-hosted GLM models cut its AI coding costs by about 90%, replacing Claude Code as default
Its ops and infrastructure lead told an AI Engineer conference that about 2,000 engineers now default to Z.ai's GLM-5.3, served on 56 Nvidia B200 GPUs at about $0.30 per million tokens.
Why it matters
Revolut is one of the clearest public cases of a large company moving everyday coding work from paid frontier tools to open-weight models it runs itself. The company says self-hosted GLM cost about a tenth of OpenAI's Codex or Anthropic's Claude Code, and that cheap tokens let engineers hand debugging and ticket work to AI. GLM-5.1 has about 754 billion parameters, with 40 billion active per token. Watch whether other large engineering teams copy the setup and how Anthropic and OpenAI respond on enterprise pricing.
Key facts
- AI coding inference costs down about 90% in under a year
- About 2,000 engineers
- Default model: Z.ai's GLM-5.3, after GLM-5.1
- Served on 56 Nvidia B200 GPUs
- Self-hosting at about $0.30 per million tokens
- Coding time improved 60%, lead time 70%