Zhipu founder Tang Jie speaks to Musk from afar: catching up with Claude Fable 5 doesn’t have to wait until 2027

Elon Musk and Tang Jie, founder and chief scientist of Zhipu, had a space conversation on X.

The reason is that a netizen asked, when can the current gap between the Chinese and American models be equalized? When do you think China will reach the Fable level? GLM-5.2 will definitely narrow the gap. ”

Subsequently, tech enthusiast Teortaxes, who can also be considered a small KOL in the tech industry (with 65K fans who always provide accurate technical interpretations of DeepSeek), responded to this.

He first assigned a position to the GLM-5.2 of Zhipu, thinking that it is currently at the level of Claude Opus 4.7-4.8. (The visual understanding part is not included, as Zhipu currently cannot achieve full modal unity. Of course, he believes that Opus has also done a poor job in this area.)

It can be inferred that there is currently a 7-month time difference between the Chinese and American models.

Based on the timeline of Mythos, he provided a reference point: the Mythos series has reached Preview level by early February 2026, with functionality equaling or surpassing Opus 4.8. If we calculate based on Claude’s catch-up speed, China would probably need to develop a “complete version of Mythos’ ability benchmark” model between November and December 2026.

Subsequently, Musk attended the event and believed that it might be delayed by another quarter, possibly in Q1 2027.

Then, the real body of GLM-5.2, who was named, arrived at the scene. Professor Tang Jie responded to Musk’s judgment by casually saying “it won’t be that long” and showing off in a low-key manner. The implied meaning is also that “domestic models represented by Zhipu, especially Zhipu, are expected to achieve breakthroughs within this year.

Industry and academic representatives in the field of AI between China and the United States shouted from afar, discussing the current issue of the “China US model gap” that everyone is most concerned about, and the comment section quickly heated up.

The onlookers began to split into two factions, with one faction being excited and believing that GLM, with its current size, has achieved remarkable results. The GLM 6 is worth looking forward to by the end of the year. For example, GLM-5.1 does not rank high enough in the global Harvey Legal Agent professional field test (reportedly zero points); But by GLM-5.2, it has already been able to break into the top three, which shows the fast iteration speed of the Zhipu model.

On the other hand, the calm faction believes that the current GLM-5.2 does not even have the ability to remember cross conversations, and it is not meaningful to just equalize on the benchmark.

Musk also agrees with this, stating that it may be possible for both China and the United States to achieve parity by the end of the year on the benchmark, but if measured by true practicality, even in Q1, it is already remarkable.

He believes that Anthropic has always focused on maximizing ‘useful intelligence’, which will not be reflected in benchmark scores, but will definitely be reflected in revenue.

This pair of benchmarks has once again pushed Zhipu’s positioning as the “Chinese version of Anthropic”.

On the one hand, both companies not only have strong academic/research oriented founding genes, but also pay more attention to underlying innovation and long-term value when pursuing technological frontiers.

On the other hand, in terms of commercialization pace and market recognition, Anthropic has broken through from the B-end coding market, occupying the minds of professional users and creating a steep growth curve and robust business model. In this regard, it is very similar to Zhipu, which also performs well in the B-end business.

Last week, Anthropic launched its latest flagship model, Claude Fable5, but was criticized for its covert model intelligence reduction and restricted access to certain regions; On the other hand, Zhipu immediately launched GLM-5.2, which topped all open-source weight models with a score of 51, significantly ahead of MiniMax-M3 (44 points), DeepSeek V4 Pro (44 points), and Kimi K2.6 (43 points), and promoted full open-source under the MIT protocol.

Comparing the attitudes of the two companies towards the open source community and users, Zhipu has not only gained a lot of reputation and popularity, but its stock price has also been on an explosive upward trend. The cumulative increase in the past five jobs has reached 99.81%, almost doubling.

In fact, Teacher Tang Jie had already made a prediction last month regarding the larger scale model update of Zhipu – native multimodal. Tang Jie, founder of Zhipu, revealed that the native multimodal model will be launched within a few months.

In early May, his answer was: go online within a few months. After Teacher Tang Jie emphasized the time rhythm again, the globalization manager of Zhipu also forwarded a tweet, predicting that “something big is about to happen”, which also means that the major version upgrade of GLM is further approaching.

After the release of GLM-5, Zhipu made efforts in Coding and Long running agent tasks, and the open source ecosystem remained at the forefront of the world. However, in terms of multimodality, especially native multimodality, it is indeed necessary to provide a clearer answer to the outside world.

How important is this answer? The K2.5 released by Kimi at the end of January this year is already a native multimodal architecture; Alibaba Qwen3.5-OMni will be launched in March, with end-to-end pre training based on over 100 million hours of audio and video data; GPT-4o completed the implementation of its native multimodal architecture in April last year.

The understanding and construction of multimodality has become the most critical dimension in bridging the gap between head models. In Tang Jie’s tweet, he shared the strategic significance of building multimodal capabilities: perceiving the environment is the foundation for completing long tasks, and multimodal is not a functional addition, but a prerequisite for agents to truly land.

Therefore, supplementing multimodality is not only a necessary condition to support the next capital narrative, but also a necessary path for Zhipu to achieve a closed-loop technological path.

But the author believes that in order to achieve the goal of catching up with Fable5, domestic models not only need to push the model to trillion level parameters in the pre training stage, but also face greater challenges in the post training stage, such as allowing the model to partially run through self training and self iteration (recursive self-improvement, RSI).

Finally, bring the topic back to the frequently discussed and constantly new topic of “the gap between the Chinese and American models”. Anthropic CEO Dario once gave his perspective on the ‘final judgment’.

In the May report “Two Scenarios for Global AI Leadership”, he made two assumptions: one is that the United States and its allies will maintain their lead, and the other is that China will catch up with the United States.

Of course, the entire report calls on the United States to lock its lead between 12 and 24 months by plugging vulnerabilities such as chip zousi, overseas data center access, and distillation attacks.

That means that by January 2028, the best model in China will at most catch up to the level of the model in the United States in January 2027, in other words, it will be one year away from starting.

But now it seems that unless Dario makes rapid progress this year, the generation gap between China and the United States is likely to further narrow.