GLM-5.3-Flash matches top models at a fraction of cost, runs without Nvidia

Z.ai's new GLM-5.3-Flash stands out not just for its performance but for its cost structure. The model reaches 57 points on the Intelligence Index, only three behind the larger GLM-5.3, while costing 0.09 dollars per task — roughly 7.5 times cheaper. This places it on the Pareto frontier of intelligence and cost, a position that adds to the recent pressure Chinese models have put on Western providers.

The infrastructure angle is equally significant. Z.ai ran the model's pre-launch traffic on Chinese AI chips, serving 100 trillion tokens a day as reported by SemiAnalysis. The company says its hardware efficiency matches common Nvidia GPUs. Because CUDA, Nvidia's programming layer, has been nearly 20 years of tuned integration, switching chips usually requires redoing that work. Z.ai instead built its own serving software on SGLang, breaking processing into independently scaling stages, which tripled throughput on the same hardware.