Summary
- Perplexity has unveiled a research preview of its fine-tuned GLM 5.2 model, designed to work within its Computer system and escalate to Claude Opus 4.8 when necessary.
- This new system operates at one-third the cost of Opus 4.8 across various benchmarks.
- This marks Perplexity’s second fine-tuning of a Chinese open-source model within 18 months, following R1-1776, which was modified to exclude around 300 topics censored by Beijing.
Perplexity has successfully adapted a Chinese open-source AI model, transforming it into a highly efficient tool that costs only a third of Claude Opus 4.8.
Today, the company announced a research preview of a modified version of Z.AI's GLM 5.2, specifically tailored for integration into its Computer agent, which is now in production.
We are excited to introduce a research preview of our new orchestrator model within Perplexity Computer.
This model is a refined version of GLM 5.2, optimized for the Computer harness. It achieves near-top performance at just 0.344 times the cost of Opus. pic.twitter.com/jcxikoFRfn
— Perplexity (@perplexity_ai) July 9, 2026
GLM 5.2, which boasts approximately 744 billion parameters, originates from Z.ai, previously known as Zhipu AI, and has been on the U.S. Entity List since January 2025. (Parameters are the various settings a model can adjust during its training phase. A higher number of parameters typically indicates a more sophisticated and capable model.) Released under an MIT license in June, this model ranks among the leading AI tools available for extended coding tasks, significantly reducing API costs.
Thanks to its open weights, users can download, modify, and commercially fine-tune it without restrictions. Perplexity leveraged this opportunity.
Understanding Fine-Tuning
Fine-tuning involves taking an already trained AI model and further training it on a more focused dataset to enhance its performance for specific tasks.
Consider it akin to customizing a car. Different mechanics can take the same Honda Civic and modify it for drag racing, enhance its appearance, or adapt it for rally driving. Similarly, in AI, developers start with a base model and adjust various parameters to enhance its expertise in a particular area, alter its biases, or adjust its constraints.
Perplexity used post-training—a method applied after the initial training—to equip GLM 5.2 with a vital skill: discerning when to execute a task independently and when to escalate to a more powerful model.
This escalation capability is central to their development. The refined GLM 5.2 features what Perplexity refers to as an "advisor tool"—a built-in function that identifies when a query is beyond its capabilities and transfers it to a third-party advanced model. Most tasks are handled without needing to consult the expensive model, saving significant costs in the process.
According to CEO Aravind Srinivas, "When paired with an advisor, this model performs at Opus 4.8 grade quality for a fraction of the cost."
We have been post-training a version of GLM that is designed to escalate to a frontier model within the Computer harness. When combined with an advisor, this model achieves Opus 4.8 grade performance at a much lower cost. Now available as a research preview! https://t.co/7y8CjOWOtI
— Aravind Srinivas (@AravSrinivas) July 9, 2026
Perplexity conducted benchmarks comparing the new system to the standard GLM 5.2 to establish a cost baseline. Utilizing the company's internal efficiency metric, which assesses the cost of completing complex tasks, they found that the fine-tuned model with an advisor is approximately twice as costly to operate compared to the basic version. However, relying solely on the top-tier Opus 4.8 model for all tasks is significantly more expensive—about 600% higher.
By integrating these tools, Perplexity's system delivers the same high-quality performance as Opus but at around one-third of the price.
Why Choose a Chinese Model and the Advantages of Open Source
The U.S.-China AI competition is often viewed as a zero-sum game. However, open-source models transcend borders. The MIT license of GLM 5.2 simplifies the process: there are no API contracts to breach, and no access restrictions that a government can impose. Users can download the weights and customize them for their needs.
Perplexity has previously navigated this path. When DeepSeek R1 gained traction in early 2025, the company adapted it into R1-1776—removing around 300 topics that were off-limits due to Chinese censorship and retraining the model to align more with U.S. perspectives. This version became a Western-hosted reasoning engine.
As Perplexity's team noted, "We cannot utilize R1's powerful reasoning capabilities without first addressing its biases and censorship."
Thus, the adaptation of GLM 5.2 follows a similar strategy, but this time the focus is on economic rather than political goals. Perplexity's Computer product already orchestrates over 19 AI models; the fine-tuned GLM aims to serve as the cost-effective default solution that manages the majority of tasks before resorting to a frontier model.
Srinivas articulated a clear long-term vision: to post-train open-source models to excel at escalation within an agent harness that already serves millions. He emphasized that Perplexity is "uniquely positioned" to achieve this due to its existing large-scale infrastructure.
The model operates on Nvidia B200 GPUs in the U.S. The next step involves a post-training of Nemotron 3 Ultra, which aims to replicate the same architecture using an American open-source model.
Comprehensive benchmarks and a research paper are anticipated in the upcoming weeks, with the model now accessible as a research preview.