Z.ai's GLM 5.2: A Cost-Efficient Chinese AI Coding Model

Z.ai, a Beijing-based lab, released GLM 5.2 on June 16 as an open-weights LLM with 753 billion parameters and 40 billion active at once, under an MIT license. Its API costs $4.40 per million output tokens, significantly cheaper than Anthropic's Opus 4.8 ($22 per million) and Fable ($44 per million). The model nearly ties Opus 4.8 on agentic coding benchmarks like FrontierSWE and PostTrainBench, and scores well on cybersecurity benchmarks, drawing comparisons to Anthropic's Mythos. However, Z.ai's own report shows GLM 5.2 trailing Opus 4.8 and OpenAI's GPT-5.5 in most coding benchmarks, such as SWE-Marathon where it completed 13% tasks versus Opus 4.8's 26%. Coders report varied experiences: Zain Hasan of Together AI says GLM 5.2 handles long-horizon tasks well and has good web design taste; David Nix of MetaRouter finds it 'really close' to frontier models for front-end work, handling 10-20% of his LLM tasks. Others like Sai Kiran Myadaram of Indhic AI had issues with rate limits and hallucinations, preferring OpenAI's Codex. The open-weights nature allows self-hosting, addressing concerns about data security for U.S. companies.