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

Perplexity released a research preview on July 9, 2026, of a post-trained version of GLM 5.2 that runs inside its Computer agent and delegates harder queries to Claude Opus 4.8. The company says the adapted model is available in production as a research preview.

GLM 5.2 is a roughly 744-billion-parameter model from Z.ai, formerly Zhipu AI, and was published under an MIT license in June. Perplexity applied post-training to the model to teach it when to handle a request itself and when to hand off to a higher-capability model inside its agent framework.

Perplexity benchmarked the adapted model against the unmodified GLM 5.2 and against using Opus 4.8 for all tasks. Using an internal efficiency metric that measures the cost to complete complex tasks, the company reported the fine-tuned GLM with an advisor runs at about 0.344x the cost of Opus 4.8. Perplexity found the post-trained GLM is roughly twice as expensive to run as the baseline GLM 5.2 but far cheaper than routing every request to Opus, which the company said is about 600% more costly.

The adapted GLM 5.2 serves as the default orchestrator inside Perplexity Computer. A built-in advisor flags queries that exceed the model’s competence and triggers escalation to the Opus model. Perplexity said most requests are handled by the cheaper orchestrator and only a subset are sent to the higher-cost model.

Perplexity stated the model runs on Nvidia B200 GPUs in the United States. The company said Computer already orchestrates more than 19 AI models and positioned the adapted GLM 5.2 as the low-cost default that handles routine work before invoking a frontier model. Perplexity also announced plans to post-train Nemotron 3 Ultra next, applying the same escalation architecture to an American open-source model.

The GLM 5.2 release follows an earlier Perplexity project that fine-tuned DeepSeek R1 into R1-1776, a version that removed roughly 300 topics the original model avoided. Perplexity noted that the MIT license of GLM 5.2 allows third parties to download and modify the weights for commercial use.

Perplexity CEO Aravind Srinivas wrote on X that the post-trained model is trained to escalate to a frontier model inside the Computer harness and that, “When paired with an advisor, this model functions at Opus 4.8 grade performance at a fraction of the cost.” The company said full benchmarks and a research paper will be published in the coming weeks and that the adapted GLM 5.2 is available now as a research preview inside Perplexity Computer.