Zhipu AI admits mistakes in GLM-5 launch, offers refunds

Chinese company Zhipu AI found itself in a tough spot after the botched launch of its GLM Coding Plan. On February 21, the company published an official letter of apology in which it not only admitted its mistakes but also announced concrete compensation for affected users. That is rare in a market where major tech players usually prefer to stay quiet about operational failures rather than air them publicly.

Zhipu AI is one of China's most prominent language model developers and the company behind the GLM model series. It has long competed with players such as Baidu and Alibaba for developers and enterprise customers. GLM-5 was supposed to be the next step forward and strengthen the platform's position. Instead, the hype around the model backfired: traffic ran far above forecasts, and the infrastructure couldn't handle the load.

In its letter, Zhipu AI openly named three key mistakes. The first was a lack of transparency in the GLM Coding Plan rules: users didn't understand exactly what they were signing up for or how access to new features would work. The second was an overly slow, phased rollout of GLM-5: the company took a cautious approach to expanding access, and combined with the hype, that led to widespread frustration. The third was a poorly designed transition path for existing customers, who were cut off from promised upgrades without a clear explanation. Notably, leadership didn't try to blame outside circumstances — all three points are framed as internal management failures.

The technical side of the problem was predictably painful. After GLM-5 shipped, traffic exceeded every expectation, and the company simply couldn't scale server capacity fast enough. It had to introduce phased access to the model: first for Max users, then Pro, and finally Lite. At the time the apology was published, Max customers already had full access. Pro users were formally connected but hit limits during peak hours because of heavy load on the clusters. Lite subscribers are still waiting their turn — the company plans to open access to them gradually during off-peak periods after the holidays. As compensation, all affected Pro and Lite users can request a refund themselves if the service didn't meet their expectations.

The story matters not just as a corporate scandal but as a symptom of a broader industry problem. The race to ship new models often outpaces infrastructure — and that isn't limited to Chinese companies. Western players have gone through similar episodes of overloaded servers and opaque access terms: think of ChatGPT's many outages at launch, or the periodic OpenAI API access limits during demand spikes. The difference is that Zhipu AI chose to act openly — publicly admitting mistakes in the tech sector is still more the exception than the rule.

For the language model market, the episode sets an important precedent. Users are getting more demanding: they expect not only high-quality models but also reliable infrastructure, clear terms, and a fast response when things go wrong. Companies that can communicate honestly with their audience during a crisis have a chance to keep users' trust even after serious failures. Zhipu AI has bet on exactly that approach — now it remains to be seen whether real fixes will back up its words.