In short
- Disruption produced an overview of the study of the GLM 5.2 model trained later, which was built to act as a musician within its computer system and access to Claude Opus 4.8 only when needed.
- This system costs a third of the price of the Opus 4.8 in benchmarks.
- This is Perplexity’s second Chinese hit in 18 months—the first was R1-1776, a version of the DeepSeek R1 that was stripped of nearly 300 heads ordered by Beijing.
The disruption has turned the Chinese open-source version into a near-margin at about a third of what Claude Opus 4.8 costs.
The company released a overview of research today it is a model trained after Z.AI’s GLM 5.2, built to work within its computing infrastructure and available in production.
GLM 5.2 is an approximately 744-billion-parameter model from Z.ai—formerly Zhipu AI, a Beijing-based lab that has been in use. US Entity List from January 2025. (Parts are all the different calls and transformations that the model can handle during training. The more advanced, more complex and more powerful model s.) Released under the MIT license in June, it is among the top AI models available at the moment for remote coding signals at an expensive API price.
Open assets mean that anyone can download, modify, and successfully modify products without restrictions. Depression did the same.
What is good planning really like?
Optimization is the process of taking an AI model that has already been trained and retraining it on small, focused data to make it better at a particular task.
Think of it like repairing a car. Different engines can be equipped with the same Honda Civic, for example, making it faster for drag racing, more stylish, more flexible for the rally, etc.
Brainstorming was used after training—the same method that is used after advanced training of the model—to teach GLM 5.2 one very important skill: knowing when to work on its own and when to go to something more powerful.
That ride is the foundation of what he built. The improved GLM 5.2 includes what Perplexity calls an “advisory tool”—one that can detect when a question exceeds its capabilities and is assigned to some type of threshold. Most jobs don’t reach the cutting edge. Only those who really need to do it.
This can save a lot of money on guesswork.
“When combined with a mentor, this model works at Opus 4.8 grade at a fraction of the cost,” CEO Aravind Srinivas wrote on X.
Disruption was performed against the standard GLM 5.2 to establish the baseline value. Using metrics that work within the company that measure the amount of money to complete complex tasks, the results showed that the improved model with a consultant is more than twice as expensive as the first model. However, using the highest quality version of Opus 4.8 for everything is very expensive (about 600% of the original price).
Combining these tools, the Perplexity system achieves the same functionality as the Opus but at about a third of the cost.
Why the Chinese brand—and why openness makes it so
US-China AI competition tends to be framed as zero-sum. In fact, opening models do not stop at borders. GLM 5.2’s MIT license makes computing easy: No API contracts to break, no access changes that the state can change. You download the weights and you can change them to whatever you want.
Paradoxes have been on this road before. When DeepSeek R1 swept the world of AI at the beginning of 2025, the company successfully placed it as R1-1776 – mapping about 300 topics that refused to discuss because of the Chinese government’s monitoring, and also trained the model to be in favor of the United States. It became a Western-hosted version of the same reasoning engine.
“We can’t use R1’s powerful logic without completely eliminating bias and censorship,” the Perplexity team wrote at the time. blog post.
So, this move to GLM 5.2 follows the same template, except this time the focus is not politics but economics. Disturbing Computer medicine already generates 19+ AI models; A well-structured GLM is designed to be a low-cost regression model that takes most of the work before touching the marginal model.
Srinivas said the long-term idea is straightforward: open-source post-rail models to improve ride-hailing, within infrastructure that already serves millions of users. Disruption is “special” to resolve, he wrotebecause the infrastructure has already been set to scale.
This version runs on Nvidia B200 GPUs in the United States. Next in line: the Nemotron 3 Ultra train, which will be based on the same architecture using an American open-source model.
The full benchmarks and research paper are expected in the coming weeks. The sample is available as a research sample.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
