GLM-5.2 has credible high-end coding evidence: 78.7% ±1.9% on Epoch’s SWE-bench Verified leaderboard, plus a 51 Artificial Analysis Intelligence Index. Its documented 1M-token context, 128K output, open weights, and MIT license are meaningful technical and deployment advantages. It exceeds the published GLM-4.6 anchor in capability, coding, speed, developer experience, and public signal on materially stronger current evidence. Coding evidence is mixed rather than uniformly leading. The independent DeepSWE v1.1 result—44% ±2%, 14th of 18 displayed models—prevents a score near GPT-5.5, GPT-5.6 Sol, or Claude Opus 4.8, all of which have substantially higher catalog coding anchors. It is broadly comparable to Kimi K2.5 on coding-oriented use, but trails its 86 coding score because the supplied independent agent benchmark is weak. No GLM-5.2 entry was found on LiveCodeBench V5. Value and serving evidence are strong: Artificial Analysis lists $1.40/M input and $4.40/M output, 115.8 output tokens/s and 1.43s median TTFT, while provider tracking reports 15 providers and lower-cost alternatives. Pricing clarity is lower than the price score because supplied official pricing excerpts omit a durable rate table. Multimodal capability remains conservatively low: the supplied materials establish a text model, not comparable image/audio I/O. Official Terminal Bench and SWE-bench Pro claims are useful but vendor-reported and not treated as independent performance proof.
