Perplexity says tuned GLM 5.2 cuts AI task costs versus Claude Opus

Perplexity says it has put a cheaper AI model into production for its Computer product, aiming to get Claude Opus 4.8-level results without sending every task to a top-priced model. For investors watching AI software, the key point is cost: inference, the compute used each time a model answers a prompt, can shape how expensive these products are to run.

The company announced on X that it released a research preview of a new “orchestrator” model inside Perplexity Computer. Perplexity said the model is an adapted version of GLM 5.2 from Z.AI, post-trained for its Computer harness, and that it delivers “near-frontier performance” at 0.344 times the cost of Opus.

In plain English, Perplexity is using one model to decide how much AI power a task needs. The tuned GLM 5.2 handles jobs it can complete on its own and passes harder ones to a stronger third-party frontier model through what Perplexity calls an advisor tool.

That matters because the most capable models are also expensive to use. If a lower-cost model can complete most work and call in a premium model only for tougher prompts, the overall cost per task can fall. Perplexity CEO Aravind Srinivas wrote on X that, when paired with an advisor, the model works at “Opus 4.8 grade performance” for a fraction of the cost.

What Perplexity changed

GLM 5.2 is a roughly 744-billion-parameter model from Z.AI, formerly known as Zhipu AI, according to Decrypt. Parameters are the learned settings inside an AI model that help determine how it processes information. Larger parameter counts can allow a model to handle more complex patterns, though size alone does not determine quality.

Z.AI is a Beijing-based lab that has been on the U.S. Entity List since January 2025, according to Decrypt. GLM 5.2 was released under an MIT license in June, which means its open weights can be downloaded, changed, fine-tuned and used commercially without the same limits attached to many closed models.

Fine-tuning means taking a model that has already been trained and training it further on a narrower set of examples so it gets better at a specific job. Perplexity used post-training, a related process applied after the main training phase, to teach GLM 5.2 when to answer directly and when to escalate a request.

Why the advisor setup matters

The advisor tool is the core of Perplexity’s setup. Instead of treating every prompt as if it needs the most advanced model available, the system is designed to sort tasks by difficulty. Easier requests stay with the adapted GLM 5.2. Requests that exceed its ability are handed off to a third-party frontier model.

Perplexity said the research preview is available now in production. The company framed the release as a way to get near-frontier results at roughly one-third the cost of Claude Opus 4.8 across benchmarks.

This is also Perplexity’s second Chinese open-source fine-tune in 18 months, according to Decrypt. Its earlier R1-1776 project was based on DeepSeek R1 and removed roughly 300 censorship topics mandated by Beijing, Decrypt reported.

The broader takeaway is that open-weight models are becoming raw material for companies that want more control over price and product behavior. Perplexity’s claim is narrow but notable: with extra training and a routing system, it says a Chinese open-source model can do much of the work while relying on Claude Opus 4.8 only when needed.

This story draws on original reporting from Decrypt.