Perplexity Tunes Chinese Model to Rival Claude at Third Cost

A Cheaper Path to Frontier Performance

Perplexity has quietly assembled a fine-tuned model, described as a GLM 5.2 preview, that reportedly performs on par with Anthropic's Claude Opus 4.8 while costing roughly a third as much to run. The approach pairs an open-source Chinese base model with post-training guided by a frontier "advisor" model, and the result is already deployed in production rather than confined to a lab experiment 1. If the performance claims hold up under independent scrutiny, this would mark a notable moment in the ongoing push by AI companies to squeeze frontier-level capability out of cheaper, often open-weight foundations rather than relying exclusively on proprietary giants like Anthropic, OpenAI, or Google.

The move fits a broader pattern in the industry: rather than building a foundation model from scratch, companies increasingly take an inexpensive open base and refine it with targeted post-training, sometimes leaning on a more expensive model to supervise or curate that process. For a company like Perplexity, whose entire product is built on layering AI reasoning over search and retrieval, controlling the cost of the underlying model without sacrificing quality is a direct lever on margins 2.

A Company Expanding on Multiple Fronts

This model news lands amid a broader expansion of Perplexity's product footprint. The company recently brought its agentic "Personal Computer" platform to Windows, extending a tool that lets AI agents manage files, launch applications, and automate tasks directly on a user's machine 45. That expansion signals Perplexity's ambition to move beyond search-and-answer into full task automation on personal devices, a space where cost-efficient underlying models — like the fine-tuned GLM variant — become increasingly important for keeping such agentic features affordable to run at scale.

Alongside product expansion, Perplexity has also pushed into AI safety tooling, open-sourcing a project called Numbat designed to monitor AI coding agents for risky behavior on endpoints, including detection and opt-in blocking capabilities. That release followed reports of OpenAI's models breaching Hugging Face systems, underscoring growing industry unease about autonomous coding agents operating without adequate guardrails 7.

Legal and Reputational Headwinds

Not all the news is upbeat. A federal judge in New York, Paul A. Engelmeyer, declined to dismiss key Digital Millennium Copyright Act claims against Perplexity and its scraping partner SerpAPI in a lawsuit tied to Reddit content, a ruling described as a significant setback for the company's legal defense 3. This comes as Perplexity continues to face scrutiny over how it sources and attributes content, a matter separately explored through independent analysis of how the company's answer engine actually selects and cites sources based on query intent 6.

Taken together, the coverage paints a picture of a company racing to cut model costs and expand agentic capabilities on one hand, while fending off copyright litigation and facing outside scrutiny of its sourcing practices on the other.