GLM-5.2 Review

The interesting number is not 753 billion parameters, it is 62.1 on SWE-bench Pro against 58.6 for GPT-5.5. An open weights model from a Chinese lab sitting at the top of a serious coding benchmark is no longer a novelty, but GLM-5.2 is the one that made the comparison uncomfortable for everybody selling reasoning by the token. Terminal-Bench 2.1 tells the other half of the story: 81.0 here against 85.0 for Claude Opus 4.8, close enough to matter and far enough to be honest about.

The licence is the real product. MIT, no regional restrictions, weights on Hugging Face and ModelScope, self hosting and fine tuning and commercial use all permitted, which is a different proposition from a cheap API whose terms can be rewritten. A one million token context window means a whole repository fits, and 128K of output means it can actually write the patch it planned instead of describing it. Sparse attention is what keeps the cost of that from turning absurd.

Two reservations. General reasoning sits a step behind the frontier, so this is a coding and agent model rather than a thinking partner, and the grades below say exactly that. And Z.ai ships fast enough that any review is a snapshot: GLM-5 in February, 5.1 in April, 5.2 in June, 5.3 in August, the last of those built on the same base with extended post training. Anyone standardising on 5.2 should expect to be a version behind within a quarter. API pricing runs around $1.40 per million input tokens and $4.40 per million output.