Home›Crypto›Market Structure›Perplexity fine-tunes GLM 5.2 to match Claude at one-t…

Perplexity fine-tunes GLM 5.2 to match Claude at one-third cost

Perplexity says its post-trained GLM 5.2 orchestrator can escalate to Claude Opus 4.8 only when needed, cutting the cost to 0.344 times Opus across benchmarks.

Reported · Written up by our desk · Market Structure

Perplexity released a research preview of a post-trained GLM 5.2 system designed to work as an orchestrator inside its Computer agent harness, escalating to Claude Opus 4.8 only when more powerful output is required, according to Decrypt.

The company said the setup delivers near-frontier performance at about one-third the price of Opus across benchmarks. Decrypt also reported that the model is already available in production.

Perplexity built the fine-tuned orchestrator from GLM 5.2, a roughly 744-billion-parameter model from Z.AI, formerly Zhipu AI, which has been on the U.S. Entity List since January 2025. Decrypt added that GLM 5.2 was released under an MIT license in June, and Perplexity used post-training to teach the model when to handle tasks itself versus when to escalate.

Decrypt said this is Perplexity’s second Chinese open-source fine-tune in 18 months, after an earlier version of DeepSeek R1 that it described as removing roughly 300 Beijing-mandated censorship topics. The publication said the system includes an “advisor tool” as part of that escalation workflow.

More like this

Sources