Z.ai's GLM-5.2 Soars Amid Anthropic's Fable 5 Ban, Redefines AI Accessibility
On June 12, the U.S. Commerce Department enacted an export control measure that barred Anthropic from supplying its Fable 5 and Mythos 5 models to foreign nationals, prompting the immediate shutdown of these systems. In response, Z.ai, previously known as Zhipu AI, introduced the GLM-5.2 model the following day under a permissive MIT license. Notably, this architecture was trained entirely on Huawei Ascend processors, entirely sidestepping Nvidia resources.
Within a week of its launch, GLM-5.2 had claimed the top position on widely recognized AI model leaderboards. Z.ai's market capitalization surged past HK$1 trillion (approximately US$128 billion), highlighting the growing influence of a company operating from the shadows of Washington's trade blacklist.
Performance Insights
GLM-5.2 exhibits a mixed bag of results, capturing first place on Design Arena's human-preference coding board with an Elo score approximately 10 points higher than Fable. It also ranks highest on Artificial Analysis's Intelligence Index v4.1, scoring 51, outpacing models from MiniMax and DeepSeek. In another test, SWE-bench Pro, it achieved a score of 62.1, outclassing GPT-5.5's score of 58.6.
However, its performance isn't consistently superior across all metrics. In assessments like Code Arena's front-end board, GLM-5.2 took second place behind Fable 5. Furthermore, on the AA-Briefcase test, which evaluates long-duration tasks, Fable 5 previously led with a score of 1,587 Elo, surpassing others, including Opus 4.8 and GLM-5.2.
GLM-5.2 continues to hold the top accessible slot since the models that previously outperformed it—Fable 5 included—are now inactive, with the latter being disabled due to the export restrictions.
Training Without Nvidia
The training stack for GLM-5.2 stands as a direct challenge to U.S. efforts aimed at stifling advancement in Chinese AI development. Z.ai has been on the U.S. Entity List since January 2025, preventing access to Nvidia's advanced GPUs. This model was trained using approximately 100,000 Huawei Ascend 910B processors and the MindSpore framework, completely avoiding any Nvidia technology.
However, the Ascend 910C offers only about 60% of the inference performance compared to Nvidia's H100, indicating a notable gap in efficiency and scalability, as highlighted in a December report by the Council on Foreign Relations. Nevertheless, there are indications from Huawei that their technology is improving; recent claims suggest that clusters of Ascend chips can now handle substantial training tasks, although they still lag behind Nvidia in chip-level performance.
The Fallout from Fable 5's Shutdown
The timing of the Fable 5 release couldn’t have been more unfortunate. Released on June 10, this model was a safety-checked version of Mythos 5 intended to mitigate potential risks. Just two days later, officials uncovered a vulnerability that could bypass Fable's safeguards, leading to its rapid deactivation for all users, including Anthropic’s non-citizen staff.
Anthropic subsequently addressed the ban, asserting that the exploit was understood to be limited in scope and reflected only known vulnerabilities similar to those found in other public models like OpenAI's GPT-5.5. The company's attempts to clarify the issue have not restored access, leaving them unable to maintain the models for domestic use only.
In contrast, GLM-5.2's permissive MIT license allows users to download and self-host its model, reinforcing its position as an accessible alternative. Yet, the operational requirements are substantial; with 744 billion parameters and a one-million-token context window, GLM-5.2 necessitates high-performance computing resources, ruling out standard desktop setups.
The most feasible way to harness GLM-5.2 is through its API, which Z.ai prices competitively at $1.40 per million input tokens and $4.40 for outputs. This pricing structure contrasts sharply with options like Opus 4.8 and Fable 5, which were significantly more costly per task.
The rapid adoption of GLM-5.2 prompted Z.ai, listed on the Hong Kong exchange as Knowledge Atlas Technology, to experience a stock increase of up to 42% by June 22, pushing its market cap above HK$1 trillion. Z.ai's founder, Tang Jie, hinted that a model comparable to Fable 5 will debut sooner than expected, setting the stage for future developments in AI technology within China.
