Perplexity is betting on cost efficiency in the AI agent space. The company has introduced a research version of a new orchestrator model for its local agent, Perplexity Computer. At its core is GLM 5.2 — an open model from China's Z.ai (formerly Zhipu AI), containing 744 billion parameters.
The key innovation is not simply replacing existing solutions, but a hybrid architecture. GLM 5.2 handles the bulk of requests, operating with performance close to cutting-edge models, but at significantly lower costs. Perplexity claims a cost of 0.344x compared to the benchmark Claude Opus. When the model encounters a task beyond its capabilities, an "advisor" tool steps in, handing control over to a more powerful model. In tandem with such an advisor, the system can demonstrate performance at the level of Opus 4.8.
The Economics of Agentic AI
Perplexity Computer is an agent system that distributes tasks among more than 19 models. The adapted GLM 5.2 becomes a cheap base layer, allowing for a substantial reduction in operational costs. The next step will be the post-training of Nemotron 3 Ultra — another open model that the company plans to adapt for its ecosystem.
It is important to note that GLM 5.2 is distributed under the MIT license, giving Perplexity complete freedom for modifications and commercial use without the restrictions typical of closed APIs. However, there is a geopolitical nuance: Z.ai has been on the U.S. Entity List since January 2025. Perplexity circumvents this restriction by hosting the adapted version of the model on its own infrastructure in the United States, fully controlling the post-processing, routing, and execution layer.
Analysis and Prospects
This move is a direct continuation of Perplexity's strategy to reduce dependence on expensive proprietary models. The company previously adapted DeepSeek R1, releasing the R1-1776 version with elements of Chinese censorship removed. In the case of GLM 5.2, the focus has shifted from political neutralization to pure economics: the goal is to create a cost-effective agent system without sacrificing quality on complex tasks.
From my perspective as an analyst: Perplexity demonstrates a mature approach to scaling AI agents. The hybrid model of a "cheap base layer + expensive advisor" is not a temporary fix, but a template for the future of the entire industry. In an environment where the inference cost of advanced models remains high, the ability to efficiently distribute workload will become a key competitive advantage. Recall that in March, Perplexity already introduced Personal Computer as a competitor to OpenClaw, operating 24/7 with preserved "memory" between sessions. The adaptation of GLM 5.2 is a logical continuation of this strategy, aimed at democratizing access to powerful AI agents.