GLM-5.2's Code Reviews Are Only as Good as Your Prompt (blog.kilo.ai)

GLM-5.2, Z.ai's recently launched open-weight model, shows promising capabilities in code review but is significantly impacted by the way prompts are framed. In a series of controlled tests, GLM-5.2 exhibited high consistency in identifying serious security flaws—catching 13 to 15 out of 16 bugs in a straightforward codebase—largely unaffected by the reasoning effort used in the review. However, when the complexity increased with subtler, multi-route bugs, the model's performance varied dramatically based on prompt wording, underscoring its strengths with local bugs but revealing limitations when broader context was needed. The findings indicate that while GLM-5.2 excels at identifying straightforward code issues, its effectiveness diminishes when bugs require holistic understanding across multiple endpoints, aligning with the behavior seen in other frontier models like GPT-5.5 and Opus 4.8. Developers are advised to carefully structure prompts to emphasize behavior and consistency, and to not rely solely on GLM-5.2 for critical changes, as unpredictability in its performance necessitates additional reviews or stronger models for comprehensive assurance.