THE ESSENTIALS IN 30 SECONDS

  • Revolut serves self-hosted GLM models to about 2,000 engineers on 56 Nvidia B200 GPUs, at a blended cost of around $0.30 per million tokens, according to a conference talk by David Abitbol.
  • The resulting invoice is about one-tenth of what the fintech previously paid Anthropic and OpenAI; frontier tools such as Claude Code and Codex remain available for specific tasks.
  • Internal metrics showed large gains in coding and lead times, but Revolut's comparison does not isolate the model change from other factors.

British fintech Revolut has shifted most of its AI-assisted coding away from metered APIs run by Anthropic and OpenAI to open-weight models it hosts itself. David Abitbol, an operations and infrastructure leader at the company, described the move in a talk at the AI Engineer conference in New York on October 13.

The setup and the savings

According to Abitbol, Revolut serves models from Zhipu AI's GLM family to roughly 2,000 engineers using 56 Nvidia B200 GPUs. The talk transcript says the company started with GLM 5.1 in April and now runs GLM 5.3, although some write-ups cite GLM-4.5 instead. The endpoints are compatible with OpenAI's interface, so engineers use them through tools such as OpenCode and Codex.

The blended cost is put at about $0.30 per million tokens, and the overall invoice at roughly a tenth of what Revolut used to pay the two US labs. StartupHub.ai points out that this per-token figure leaves out cloud spending, staff and observability costs. Self-hosted models are now the default, while Claude Code, Codex and Gemini remain available where they clearly perform better. Image tasks are still sent to hosted multimodal models, because the self-hosted model handles only text.

Abitbol said default inference settings performed poorly at first. Tuning parameters such as context length, quantization and parallelism, without adding hardware, raised throughput from about 3 to more than 7 requests per second and cut time to first token from 14.26 seconds to 1.68 seconds.

Productivity claims and caveats

Revolut measured coding time and lead time from picking up a feature to merge-ready code. Early use of Gemini brought gains of roughly 10–20%, while Claude Code under strict per-engineer budgets showed little effect at first. After the switch to unrestricted self-hosted models, Abitbol reported improvements of about 60% in coding time and 70% in lead time. The comparison does not separate the model change from other factors, and the results have not been independently verified.

On security, the company says each agent session runs in a sandbox with a human in the loop and command allow-lists. Model weights are run on infrastructure in jurisdictions with data-equivalency arrangements (the US, EU, UK and Israel), with zero data retention by the inference provider, and no requests go to Chinese endpoints. The system covers the internal coding tools, not customer-facing banking systems.