GLM 5.2

Z.ai announced GLM-5.2 on June 16, 2026 as its flagship model built for long-horizon tasks, introducing a solid 1-million-token context window designed to remain reliable across extended coding-agent trajectories. The release emphasizes that a long context must be engineering-usable, not merely wide, and the model ship Architecturally, GLM-5.2 introduces IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at the 1M context length, along with an upgraded Multi-Token Prediction (MTP) layer that increases speculative-decoding acceptance length by up to 20%. The model is r

Z.ai officially announced GLM-5.2 on June 17, 2026 (Japan time) as a Mixture-of-Experts model with roughly 753B total parameters but only about 40B active per inference, balancing scale with lighter compute. The weights ship as an open model under the MIT License on Hugging Face at zai-org/GLM-5.2, letting teams download, fine-tune, and self-host the model. GLM-5.2 is reported to outperform GPT-5.5 on several software-development benchmarks and rank 2nd on a coding leaderboard at launch. An efficiency mechanism called IndexShare reuses computation across sparse attention layers to cut per-token cost over the 1M-token context window, targeting long-horizon coding and agent workloads.

VentureBeat reported on June 16, 2026 that Z.ai (formerly Zhipu AI) released GLM-5.2, a 753-billion-parameter open-weights LLM targeted at long-horizon autonomous coding and engineering work, immediately available on Hugging Face, the Z.ai API, and more than 20 third-party coding environments. The model pairs a highly The report confirms the IndexShare optimization (one indexer reused across every four sparse attention layers, cutting per-token FLOPs by 2.9× at 1M context) and the upgraded Multi-Token Prediction layer for speculative decoding (up to 20% longer accepted token length). It also cites Z.ai's claim that GLM-5.2 beats GPT

LLM-Stats ranks GLM-5.2 at #40 overall on its composite score, with 397 votes showing stable quality and 53 points lower real-task performance than the model's self-reported results. Across conversation depth, the model's score drifts from 15.8 at turn 1 to 13.5 by turns 31+, indicating meaningful degradation on very long agent-style conversations. GLM-5.2 is listed at a blended price of $0.83 per 1M tokens, sitting between DeepSeek-V4-Flash-0731 at $0.066 and Claude Opus 5.5 at $4.76 on LLM-Stats' cost-efficiency curve. In head-to-head capability versus price, it scores 44.9 against Claude Opus 5.5's 60.5, positioning it as a mid-cost long-context open-weight option.

NIST's Center for AI Standards and Innovation (CAISI) released an independent assessment of Z.ai's GLM-5.2 on July 8, 2026, finding it was probably the most capable open-weight model at its June 16, 2026 release. CAISI's evaluations place GLM-5.2's overall capabilities at a level similar to GPT-5.2 (December 2025) and its cyber capabilities similar to Opus 4.6 (February 2026). CAISI's safeguard assessment of GLM-5.2 is mixed: its safeguards allow assistance with agentic cyber exploit development and block fewer sensitive biological questions than reference U.S. models, yet the model appears potentially more robust against agent hijacking and prompt-based jailbreaks than other evaluated PRC open-weight models. The report notes open-weight safeguards can be circumvented when self-hosted.

BenchLM catalogs GLM-5.2 with a capability score of 61.5/100, ranking it 47th of 216 tracked models, and explicitly marks the release as superseded by Z.ai's newer GLM-5.3. Coding is the model's strongest published category, ranking 31st of 146, while knowledge ranks 49th and agentic ranks 33rd with 26 verified benchmark rows. GLM-5.2 is reported with a 1M-token context window, 95 tokens/second throughput, a 23.99-second first-token latency, and pricing of $1.40 input and $4.40 output per 1M tokens. The supersession note is the clearest signal that GLM-5.2 is no longer Z.ai's current frontier offering, even though its benchmark footprint remains substantial.

Z.ai's GLM-5.2 launched on June 16, 2026 as an open-weight model, 70 days after GLM-5.1, with 744B total parameters and a 1M-token context window. Reasoning/thinking mode is supported, and benchmark coverage spans BullshitBench v2, ProgramBench, SWE-Bench Pro, DeepSWE 1.1, Next.js Evals, and Frontier-Bench v0.1. Provider pricing is verified against docs.z.ai as of August 18, 2026 at $1.40 per 1M input tokens, $0.26 per 1M cached input, and $4.40 per 1M output tokens, all in USD. The release gives engineering teams a long-context open-weight option for code generation and agent tasks with the same pricing tier across Z.ai's own endpoints.

Z.ai released GLM-5.2 on June 16, 2026 as an open-weight flagship foundation model targeting long-horizon reasoning and coding workloads. The release is documented with a verified 1M-token context window and a 128K maximum output, positioning the model for repo-level coding, multi-step software engineering, and sustained agent execution over very large documents. Z.ai's official pricing sets GLM-5.2 at $1.40 per 1M input tokens, $4.40 per 1M output tokens, and $0.26 per 1M cached input tokens, with cached input storage offered as limited-time free. The model is distributed under an MIT license with public weights, making it directly downloadable for self-hosted fine-tuning and evaluation.