Z.ai Releases GLM-5.2: The 753-Billion Open-Weight Model Challenging GPT-5.5 at One-Sixth the Cost

In a move that reshapes the competitive landscape of artificial intelligence, Chinese startup Z.ai, formerly Zhipu AI, has released GLM-5.2, a 753-billion parameter open-weights large language model under the permissive MIT license. Designed to excel in long-horizon autonomous coding and engineering tasks, the model surpasses leading proprietary models like OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 on several benchmarks, while costing a fraction of the price.

Available immediately on Hugging Face and through the Z.ai API, GLM-5.2 features a highly stable 1-million-token context window and enterprise subscription tiers starting at just $12.60 per month. The MIT license allows enterprises to download, customize, and run the model locally or on virtual machines, paying only for compute and electricity. This is especially attractive given recent U.S. regulatory uncertainty: the Trump Administration's export control directive last week restricted foreign nationals from using Anthropic's Claude Fable 5, prompting Anthropic to take the model offline entirely.

Architecturally, GLM-5.2 introduces IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token compute FLOPs by 2.9x at maximum context length. It also incorporates an upgraded Multi-Token Prediction (MTP) layer for speculative decoding, boosting accepted token length by up to 20%. Selectable thinking modes, "Max" for peak intelligence and "High" for latency-sensitive efficiency, allow developers to balance performance and cost.

Benchmark results are impressive: GLM-5.2 scores 62.1% on SWE-bench Pro, beating GPT-5.5 (58.6%) and its predecessor GLM-5.1 (58.4%). On FrontierSWE, it achieves 74.4%, surpassing GPT-5.5 (72.6%) and nearly tying Claude Opus 4.8 (75.1%). MCP-Atlas yields 77.0, ahead of GPT-5.5 (75.3). In Humanity's Last Exam with tools, it scores 54.7 vs. GPT-5.5's 52.2. On extended workloads like PostTrainBench and SWE-Marathon, it leads with 34.3% vs. 25.0% and 13.0% vs. 12.0% respectively. While slightly trailing on Terminal-Bench 2.1 (81.0 vs. 85.0), it outperforms Google Gemini 3.1 Pro (74.0) and takes first place on Design Arena with an ELO of 1360, besting even Claude Fable 5.

The GLM Coding Plan targets developer workflows: Lite at $12.60/month for small repos, Pro at $50.40 for medium projects, Max at $112 for heavy workloads. API pricing is $1.40 per million input tokens and $4.40 per million output tokens, with a cached input rate of just $0.26. In contrast, GPT-5.5 output costs $30 per million tokens, and Claude Opus 4.8 costs $25. As AI observer Lisan al Gaib noted on X, "frontier labs are absolutely scamming you on API pricing." The MIT license, without regional restrictions, offers a path to avoid vendor lock-in and geopolitical risks. The developer community has embraced GLM-5.2 immediately; tools like Kilo Code, Cline IDE, and Eigen AI have integrated it within hours. For more on regulatory impacts, see the related article on Anthropic's feud with the Trump administration. For a broader overview of large language models, visit the Wikipedia page.

Source: https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost