Revolut’s AI cost reduction hits 90% with Chinese GLM models

What it claims

Revolut's head of AI operations, David Abitbol, says the company now spends about one-tenth of its former OpenAI and Anthropic bill. It serves about 2,000 engineers from Z.ai's GLM 5.3 on 56 NVIDIA B200 GPUs. The move is not total: Claude Code, Codex and Gemini remain available for tasks where they clearly beat the self-hosted model. Revolut had earlier put spending limits on Claude Code because heavy usage made it expensive.

How it works

Revolut downloads the GLM 5.3 weights and runs them on infrastructure it controls, so inference requests do not go to Z.ai. The serving stack is tuned for low first-token latency and many concurrent requests. Abitbol names these levers:

• FP4 quantization. He says FP8 would need about 40% more GPU resources, with no meaningful accuracy gain for Revolut's developer workloads.
• KV cache reuse. The target is a cache hit rate above 90% for programming tasks, using recurring code patterns to cut latency and GPU load.
• Context limits. Long contexts consume memory and reduce concurrency. Revolut trims context when extra context is not needed, even though some models accept up to a million tokens.
• Further options. Speculative decoding and optimization of request-processing phases are under consideration. The page does not say whether either is in use.

Coding agents run inside sandboxes, under authorization policies that allow or block agent operations, with human oversight. Telemetry records activity so that agent actions can be reconstructed and anomalies found.

Evidence

All figures are Revolut's own, as reported by Abitbol:

• Latency. Time to first token fell from 14.26 s to 1.68 s (88%) between April and August 2026. GPU count stayed at 56 while requests per second rose 76%.
• Cost. The average is $0.30 per million tokens.
• Developer metrics. Coding speed is up 60% and development lead time is down 70%. The page does not state the baseline for either figure.
• Earlier deployment. A Gemini rollout in January 2025 showed coding time down 20% and lead time down 10% in early internal data.

Abitbol's operating targets are TTFT under 3 seconds, with responses feeling slow past 5 seconds. He considers about 200 tokens per second of generation enough for engineering work, because developers increasingly review agent output after the fact.

Limits

The source is a news article, produced with AI assistance and reviewed by editors, that relays one executive's statements. It names no inference engine, serving framework, batch configuration or GLM 5.3 model size. It gives no licence terms for the weights and no accuracy evaluation behind the FP4 claim. The 90% saving compares against a prior API spend that the article never quantifies. It is unclear whether the $0.30 figure includes hardware amortization. The productivity numbers come with no method or measurement baseline.