Chinese AI lab Z.ai — formerly known as Zhipu, one of the country's so-called "six tigers" of artificial intelligence — released GLM-5.2 in mid-June 2026, an MIT-licensed open-weights model that matches or surpasses OpenAI's GPT-5.5 on several long-horizon coding benchmarks while costing roughly one-sixth as much to run. The launch became one of the most discussed AI stories of the month, with researchers describing it as a genuine step change for open models.
Key Highlights
- GLM-5.2 is a Mixture-of-Experts model with approximately 753 billion total parameters and about 40 billion active per token, built for agentic software engineering rather than chat.
- It ships under a permissive MIT license, with weights available on Hugging Face and ModelScope and an OpenAI- and Anthropic-compatible API.
- A one-million-token context window is available opt-in, alongside selectable reasoning effort modes.
- API pricing lands near $1.40 per million input tokens and $4.40 per million output tokens — roughly one-sixth of Fable 5's $10 and $50 rates.
Details
GLM-5.2 is engineered for long, multi-step agentic coding work rather than conversational chat. Its Mixture-of-Experts design activates only about 40 billion of its roughly 753 billion parameters per token, keeping inference costs low. A new sparse-attention scheme that Z.ai calls IndexShare reuses attention indexers across layers, which the company says cuts per-token compute by about 2.9 times at a one-million-token context length. The model also uses multi-token prediction to speed up generation.
On benchmarks, GLM-5.2 posts a SWE-bench Pro score in the low 60s and a Terminal-Bench 2.1 figure around 81, with a FrontierSWE result reported near 74 percent. On several long-horizon coding tasks it edges past GPT-5.5 while landing just below Anthropic's Claude Opus 4.8. On the community-run Design Arena, early testers found it ranking first, ahead of Claude Fable on front-end design tasks.
Impact
The combination of frontier-class coding scores, an unrestricted MIT license, and pricing at a fraction of closed competitors is what made the release land so hard. An open model that beats GPT-5.5 on real software-engineering benchmarks at roughly a sixth of the cost was, as one analyst put it, "not supposed to exist yet." For developers and enterprises, it means a credible, self-hostable alternative to closed frontier models for agentic coding pipelines.
The economics are significant for the MENA region. An MIT-licensed model with no regional restrictions can be deployed on private infrastructure, helping organizations meet data-residency and sovereignty requirements under frameworks such as Tunisia's INPDP and Saudi Arabia's PDPL without sending source code or proprietary data to a third-party API.
Background
Z.ai's release continues a pattern in which Chinese open labs trail U.S. closed frontier models by roughly six to nine months while shipping weights publicly. Observers compared the moment to DeepSeek R1's earlier demonstration that open labs could replicate advanced reasoning, noting that GLM-5.2 goes further by performing credibly as a general agent inside popular coding harnesses such as Claude Code, OpenClaw, and Cline.
What's Next
Independent evaluation remains the open question. Analysts caution that several circulating benchmark figures are vendor-reported or early, and advise teams to test GLM-5.2 on their own repositories before committing. Z.ai has signaled even higher ambitions, with commentary suggesting the lab is forecasting an "Open Fable"-class release by the end of the year — a marker of how quickly the open-weights frontier is closing the gap with proprietary leaders.
Source: VentureBeat