GLM 5.2 Clearly Explained (and how to set it up)

GLM 5.2 is an open-source local AI model from Zed AI that some in the community are calling the ChatGPT moment for local AI. It scores 81 points on Terminal Bench 2.1, placing it roughly four points behind Claude Opus 4.8 on that benchmark, and achieves 62.1 percent on long horizon task evaluation compared to Opus at 69.2 percent. It performs particularly well on front-end execution based tasks. The host, Amir, skipped version 5.1 and went straight to 5.2, and he acknowledges he does not fully understand what benchmark scores mean in practice, preferring live testing over relying on numbers alone.

The cost difference between GLM 5.2 and Opus 4.8 is the most immediately practical argument for the model. At approximately 50,000 input tokens and 85,000 output tokens, GLM 5.2 costs 44 cents versus $2.38 for Opus 4.8, roughly a five times price difference. That gap compounds quickly when models run continuously on large tasks. Many companies that mandated AI adoption in their first year are now concerned about excessive token spending, including non-engineering teams using high-end models like Opus 4.8 for simple tasks such as formatting emails.

GLM 5.2 can be run locally if the hardware supports it or accessed through open model providers such as Open Router. To use it in Cursor, the setup involves pasting the Zed AI API key into the OpenAI field, overriding the OpenAI endpoint with the GLM API endpoint, and adding GLM 5.2 as a custom model. Open Router provides its own API key and endpoint as an alternative, and GLM 5.2 can be configured in Codex by creating a profile with the model details and context window. Open Router is credit-based and costs around twenty dollars to start. The model carries a 1 million token context window.

GLM 5.2 does not support vision capabilities, which limits its ability to process screenshots directly. A workaround is to use Opus 4.8 to import and describe an image and then pass that description to GLM 5.2 to act on. More broadly, local models including GLM 5.2 still have significant work to do on tool capabilities and visual modalities. The model is also resource intensive and many existing consumer computers may not be able to run it locally, which constrains who can benefit from the cost savings without paying for API access.

Amir's recommended approach is to use GLM 5.2 for tasks it handles well and call Opus or Codex for others. He frames the correct mindset as token minimizing combined with output maximizing rather than token maximizing. One example workflow is to plan with Opus, review and execute with GLM 5.2, then review again with Composer 2.5 or Codex 5.5. Open Router supports sequencing between two models in what it calls a fusion or compounding model approach. Model-agnostic tools like Cursor that can run multiple models across the same task sequence stand to benefit most from lower-cost models like GLM 5.2 slotted into the right stages.

The speakers predict that VC-backed token subsidies will eventually decrease as AI companies mature and go public, drawing a comparison to how Uber subsidized rides early on before pricing normalized. Within a year some companies may shift to running local models on their own machines due to rising token costs, and investing in local compute hardware now could pay off as future models like GLM 5.3 or 5.5 emerge with improved capabilities. The broader implication is that the current window of cheap API access may not last, and teams that build workflows around cost-efficient models now will be better positioned when subsidy-driven pricing ends.