GLM 5.2 from Z.ai has emerged as the first open-weights model to genuinely compete with OpenAI's GPT and Anthropic's Claude Opus, according to a detailed analysis by technology commentator Martin Alderson.

Alderson, who has been testing the model for several weeks, describes it as "genuinely very good and hard for me to tell the difference between Opus." The model represents a significant shift in AI economics, offering what he calculates as more than 50% cost savings for most workflows.

Pricing disrupts frontier model economics

GLM 5.2 is priced at approximately $4.40 per million tokens through providers like Z.ai and Fireworks. This represents less than 20% of Claude Opus's retail price and around 15% of GPT 5.5's cost.

The model's competitive positioning comes despite some notable limitations. Alderson highlights slower inference speeds due to extensive reasoning processes, lack of vision capabilities, and poor web search integration as current weaknesses.

"It's genuinely frustrating it not being able to read image-based PDFs, screenshots and design files," Alderson writes, noting how quickly vision capabilities have become essential for daily AI workflows.

Migration barriers remain low

Both Z.ai and Fireworks offer OpenAI and Anthropic-compatible endpoints, making switching "absolutely trivial" according to Alderson. Users can simply change the base URL and API key to migrate existing applications.

This contrasts sharply with traditional enterprise software migrations that require years of planning. "The switching costs are incredibly low," Alderson notes, particularly as frontier labs frequently change policies and terms.

Enterprise adoption faces hurdles around data privacy, particularly with Z.ai's mainland China connections. However, the open-weights nature allows deployment through alternative providers or on-premises hosting.

Infrastructure costs declining

Wafer's recent analysis suggests running GLM 5.2 inference on AMD hardware costs 2.75x less per token than Nvidia Blackwell systems. Alderson expects serving costs to decrease further as optimization improves.

The analysis positions this as the beginning of significant margin compression in AI inference, with implications for the broader industry structure that Alderson promises to explore in a forthcoming second part.