GLM-5.2 released: Z.ai's 753B open-weight model matches GPT-5.2, draws a federal safety assessment

Z.ai released GLM-5.2 in mid-June, bringing a powerful glm 5.2 coding candidate with roughly 744 billion total parameters (about 40 billion active per token) and a 1-million-token context window. Weights shipped under the MIT license alongside a metered API, making it commercially usable and self-hostable from day one.

The unusual part of the launch was what Z.ai didn't publish: no official benchmark numbers accompanied the release, which is rare for a flagship-scale model. Independent evaluation has since filled the gap. Artificial Analysis puts GLM-5.2 (max) at 51 on its Intelligence Index, well above the median for open-weight models of similar size, and Z.ai's own model card — released after launch — reports 62.1% on SWE-bench Pro and 81.0% on Terminal-Bench 2.1 using Anthropic's Terminus-2 harness. The model leads open-weight competitors on public coding and agentic benchmarks, though it doesn't surpass frontier closed models like Claude or GPT.

Pricing has fallen sharply since release — from roughly $1.40 per million input tokens at launch to as low as $0.93 today via some providers, with output around $3-4.40 per million tokens depending on host. GLM-5.2 is designed for long-horizon, repository-scale coding work and runs inside agent harnesses like Claude Code and Cline via the GLM Coding Plan.

The model continues Z.ai's pattern from GLM-5.1: prioritize open weights and long-context agentic coding over frontier general intelligence. With no official benchmarks at launch and vendor-reported architecture details still unverified by a full technical report, treat the strongest claims as provisional until independent testing catches up.