China’s GLM‑4.5: Cheaper, leaner, more “agentic” AI
China’s GLM‑4.5: Cheaper, leaner, and more “agentic” At the 2025 World AI Conference in Shanghai, Chinese AI startup Z.ai (formerly Zhipu) unveiled its new GLM‑4.5 series, a family of open‑source language models designed to undercut rivals on price and efficiency. The headline model runs on just eight Nvidia H20 chips—half the hardware used by competitor DeepSeek’s R1—and Z.ai says it will charge only 11 ¢ per million input tokens and 28 ¢ per million output tokens via its API. In other words, the cost of generating text with GLM‑4.5 is more than ten times lower than DeepSeek’s current rates. Z.ai also touts an “agent‑native” design, meaning GLM‑4.5 can break complex instructions into subtasks and handle multi‑step reasoning. The company is releasing three sizes: the full GLM‑4.5 with 355 billion parameters, a streamlined “Air” version that activates only a small subset of parameters for efficiency, and a “Fl...