Zhipu AI acknowledged mistakes in the GLM-5 launch and offered compensation

Zhipu AI Acknowledges Mistakes in GLM-5 Launch and Offers Compensation

Chinese company Zhipu AI found itself in a difficult position following a failed start to the GLM Coding Plan program. On February 21, the company published an official apology letter in which it not only acknowledged the mistakes made but also announced specific compensation measures for affected users. This is a rare case on the market, where major technology players typically prefer to downplay operational failures rather than make them public.

Zhipu AI is one of the most prominent developers of language models in China, behind the GLM model series. The company has long competed with players such as Baidu and Alibaba in the battle for developer and enterprise client attention. The launch of GLM-5 was supposed to be another step forward and strengthen the platform's position, however the hype around the model turned against the company itself: traffic turned out to be significantly higher than projected, and the infrastructure could not handle the load.

In its letter, Zhipu AI openly identified three key mistakes. First — insufficient transparency of the GLM Coding Plan program rules: users did not understand what exactly they were signing up for and what the conditions for access to new features were. Second — too slow phased rollout of GLM-5: the company chose a cautious approach to expanding access, which combined with the hype led to widespread dissatisfaction. Third — a carelessly designed transition mechanism for existing customers, who found themselves cut off from promised updates without a clear explanation of why. Notably, management did not attempt to shift responsibility to external circumstances — all three points are formulated as internal management failures.

The technical side of the problem turned out to be predictably painful. After the release of GLM-5, traffic exceeded all expectations, and the company simply did not manage to increase server capacity at the required pace. As a result, they had to introduce phased access to the model: first for Max-tier users, then for Pro, and finally for Lite.

At the time of the apology's publication, Max-tier clients had already received full access, Pro users were formally connected, but during peak hours faced limitations due to high cluster loads. Lite-tier subscribers await their turn — the company plans to gradually open access to them during off-peak periods after the holidays. As compensation, all affected Pro and Lite-tier users were given the opportunity to independently request a refund if the service did not meet their expectations.

This story is important not only as a corporate scandal but also as a symptom of a broader industry problem. The race to release new models often outpaces infrastructure capabilities — and this is not limited to Chinese companies. Similar situations with server overloads and opaque access conditions have been experienced by Western players: recall the numerous ChatGPT outages at launch or periodic limitations on OpenAI API access during periods of high demand. The difference is that Zhipu AI decided to act openly — public acknowledgment of mistakes in the technology sector remains more of an exception than a norm.

For the language model market, this episode sets an important precedent. Users are becoming more demanding: they expect not only quality models but also reliable infrastructure, clear terms, and prompt responses to problems. Companies that know how to communicate honestly with their audience during crises have a chance to maintain trust even after serious failures. Zhipu AI has bet on such an approach — now it remains to see whether real actions to fix the situation will back up these words.