GLM-5.2, an open-source AI model from China's Z.ai, now beats GPT-5.5 on coding benchmarks at one-sixth the cost

GLM-5.2, an open-source large language model from Beijing-based Z.ai, is now the highest-scoring openly available model on the Artificial Analysis Intelligence Index. Z.ai is pricing it at roughly one-sixth of GPT-5.5 and one-tenth of Anthropic's Claude Fable 5, the proprietary models it matches or beats on the coding and agentic benchmarks developers actually use.

The model is a 753-billion-parameter open-weight release under an MIT license, hosted on [Hugging Face as zai-org/GLM-5.2](https://huggingface.co/zai-org/GLM-5.2), with a 1-million-token context window tuned for long-horizon coding and agentic work. On the Artificial Analysis leaderboard it scores 51, the top position among open-weight models. VentureBeat reports it edges GPT-5.5 62.1 to 58.6 on SWE-bench Pro while sitting just below Anthropic's Opus 4.8 at 69.2, and posts 74.4 percent on FrontierSWE against 72.6 for GPT-5.5 and 75.1 for Opus 4.8. On those numbers, the price gap, not the benchmark gap, is the more actionable difference for most teams.

The release landed on June 13, 2026, one day after a US Commerce Department action reported by memeburn forced Anthropic to disable Claude Fable 5 and Mythos 5 worldwide on national-security grounds. That timing is the news peg: a high-capability open-weight model appearing the day a proprietary rival goes dark. Much of the coverage has treated it as a US-versus-China story. The factual basis for the Fable 5 shutdown is currently a single outlet's reporting, and the export-control narrative around those model names is unusual enough that the timing should be read as suggestive rather than settled.

Z.ai priced GLM-5.2 at the Artificial Analysis-quoted ratios and shipped ZCode, a free desktop coding environment built around the model, aimed at Cursor, Claude Code, and GitHub Copilot. The combination of open weights, an MIT license, a million-token window, and a free IDE collapses the cost barrier that kept frontier AI behind expensive API subscriptions and gives teams a model they can run, fine-tune on their own codebase, and audit end-to-end.

Z.ai founder Tang Jie, in remarks carried by the memeburn explainer, framed the release around openness, saying "intelligence should be open." That is founder color rather than third-party validation, and it sits inside an explainer whose headline leans into a "Silicon Valley worried" frame that the underlying benchmark evidence doesn't require. The substantive critique of the model is narrower and more technical.

Independent security firm Graphistry ran GLM-5.2 through CyberBT-CTF, a capture-the-flag benchmark for offensive cybersecurity tasks, and called it the strongest open-source model they have tested, though still behind closed frontier systems. Graphistry flagged the same access story that helps developers as a dual-use risk: an MIT-licensed, top-tier open-weight model can be downloaded, fine-tuned, and deployed by anyone, including actors building offensive security tooling. That concern lands on the open-weights debate rather than on Z.ai specifically, but it is the most concrete reason the launch has drawn security attention.

Open-weight models from Chinese labs are pulling meaningful share among developers in markets US export controls reach only loosely, according to RAND's research report on global LLM usage. GLM-5.2 is the highest-scoring example of that pattern to date, and the next concrete data points are the Artificial Analysis re-score within the next month, the first independent security disclosures against deployed GLM-5.2 instances, and whether Anthropic brings Fable 5 back online or replaces it with a successor that doesn't run afoul of the same US Commerce action.