Perplexity released a research preview of a model it post-trained from Chinese open-source artificial intelligence (AI) model GLM 5.2 to fit its agent system. The model delivers performance comparable to Anthropic's Claude Opus 4.8 at about one-third of the cost.

On July 9, blockchain outlet Decrypt reported that the model is designed to coordinate tasks within Perplexity's computer environment and hand them off to Claude Opus 4.8 only when needed.

Perplexity post-trained Z.ai's GLM 5.2 for its "Computer" agent environment. The model is currently offered in a production environment as a research preview. Perplexity explained that the cost is about 0.344 times that of Opus while delivering performance close to frontier-level.

The core is a structure that does not use a high-cost model for every request. Perplexity applied an advisor tool to GLM 5.2 to distinguish tasks it can handle itself from those it should pass to a stronger external model. GLM 5.2 handles most tasks, and only some high-difficulty tasks are transferred to Claude Opus 4.8.

Cost comparisons also reflected the structure. Based on Perplexity's internal efficiency metrics, the post-trained model with the advisor cost about twice as much to operate as the base GLM 5.2. Still, compared with processing all tasks with Opus 4.8, the cost was about 600 percent lower.

GLM 5.2 is a model with about 744 billion parameters that Z.ai released under an MIT license in June. It provides open weights, allowing anyone to download it, modify it or post-train it commercially. Perplexity said it used that structure to tailor the model to its own purposes.

The release is Perplexity's second post-training case involving a Chinese open-source model in the past 18 months. The company previously released "R1-1776" based on DeepSeek R1. The new model runs on Nvidia B200 GPUs in the United States, and Perplexity plans to release full benchmarks and a research paper in the coming weeks.

The key point of the release is that it puts forward an operating structure that uses an open-source model as a low-cost default processing model and hands off to a high-performance model only when necessary. Perplexity is seeking to improve the cost efficiency of its agent system by repeatedly applying the same type of post-training.

We're releasing a research preview of a new orchestrator model in Perplexity Computer. The model is an adapted version of GLM 5.2, post-trained for the Computer harness. It delivers near-frontier performance at 0.344x of the cost of Opus. pic.twitter.com/jcxikoFRfn