The open-weight model from Zhipu AI matches Claude Opus 4.8 on quality while cutting per-task costs by more than a third
An open-weight model just crashed the enterprise coding party, and the incumbents should probably be nervous. Databricks ran GLM-5.2 through a multi-million-line internal test in July 2026 and found that it performs on par with Claude Opus 4.8, one of the top closed models, while charging significantly less per task.
The cost difference is not subtle. GLM-5.2 came in at $1.28 per task compared to $1.94 for Claude Opus 4.8. That’s roughly 34% cheaper for equivalent quality, which is the kind of math that makes enterprise procurement teams sit up very straight in their chairs.
The benchmarks tell the story #
GLM-5.2, developed by Chinese AI lab Zhipu AI (also known as Z.ai), was released around mid-June 2026 under an MIT license. That licensing detail matters: it means companies can deploy, modify, and build on the model without the kind of vendor lock-in that comes with proprietary systems.
On coding benchmarks specifically, the model is leading the open-weight pack. It scored 62.1% on SWE-bench Pro, a benchmark that tests real-world software engineering capabilities. On Terminal-Bench 2.1, it posted an 81.0% score.
The architecture under the hood is equally noteworthy. GLM-5.2 uses a Mixture-of-Experts design with approximately 744 billion total parameters, but only about 40 billion are active during any given inference. It also features a 1 million token context window, which is the kind of capacity that lets the model ingest and reason over entire codebases rather than just isolated files.
Speed and security add to the case #
Databricks didn’t just test quality. They also optimized inference speed, achieving up to 392 tokens per second for GLM-5.2.
In independent evaluations using Semgrep, a popular static analysis tool, GLM-5.2 outperformed several configurations of Claude in cybersecurity-related coding tasks. Pricing for security-focused work came in at around $0.17 per vulnerability, which is substantially lower than what closed alternatives charge.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our