A Chinese AI lab has posted benchmark results that challenge the idea that bigger models are always better. Z.AI's GLM-5.2 scored higher than GPT-5.5 on standard coding tests while costing roughly one-sixth as much to run.
The benchmark results
Z.AI shared internal evaluations this week showing GLM-5.2 outperforming GPT-5.5 across multiple coding benchmarks. The tests measure a model's ability to generate correct code from natural-language prompts, solve algorithm challenges, and fix bugs. GLM-5.2 posted higher accuracy scores on tasks like HumanEval and MBPP, two widely used coding metrics.
The company did not release raw numbers, but the performance gap was large enough to make the cost comparison the bigger story. GPT-5.5, by contrast, is the latest iteration of a series that has dominated the AI race for months.
The cost advantage
Pricing data from Z.AI shows GLM-5.2 costs about one-sixth what it takes to run GPT-5.5 on equivalent workloads. That means a company doing 100,000 code-related queries a day could see annual savings in the millions of dollars by switching.
The savings come from a combination of smaller model size and more efficient architecture. GLM-5.2 uses a sparse mixture-of-experts design that activates only a fraction of its parameters for each task. GPT-5.5 is believed to be a dense model — though exact specs are not public — requiring all parameters for every request. That difference directly drives down server costs for Z.AI and, in turn, for customers.
How the competition shifts
Z.AI has been a quieter player in the global AI arms race, but this release puts it on a collision course with the makers of GPT-5.5. If GLM-5.2 can maintain its lead on real-world developer tasks — not just curated benchmarks — it could force competitors to cut prices or speed up their next model launches.
Enterprise buyers are increasingly focused on cost per output. A model that delivers better code for a fraction of the price changes the math for startups and large companies alike. Z.AI has not disclosed a general release date for GLM-5.2, but early access partners are already testing it.
For now, the benchmark numbers are Z.A.I's word against the market's. Independent verification from a third party like Stanford's Center for Research on Foundation Models would carry more weight. The company says it will publish full evaluation details in the coming weeks.
