Chinese AI startup Z.ai issued a rare public apology on February 21 after its new large model, GLM-5, rollout triggered user backlash over billing, throttling, and flaws in its upgrade mechanism, underscoring operational risks as large AI models rapidly advance.
The company said GLM-5's actual compute consumption exceeded that of previous versions, prompting Z.ai to set peak and off-peak billing multipliers at 3x and 2x, respectively, without full prior disclosure. Users reported abnormal usage patterns and rapid cost increases, raising concerns about the charging system's transparency.
Traffic surged far beyond expectations after launch, Z.ai acknowledged, with black-market and scalper actors deliberately consuming compute resources. Capacity expansion lagged behind demand, forcing staged throttling by user tier. Some paying customers temporarily lost access to core model services, damaging user experience and prompting complaints.
Operational failings and user impact
Z.ai also conceded that its upgrade mechanism lacked safeguards for existing users. Some customers who mistakenly upgraded plans suffered rights losses, and the company failed to adequately disclose these risks or provide real-time feedback channels, sparking controversy. In response to the backlash, Z.ai announced a package of corrective and compensation measures.
The announced measures include publishing full billing rules, shortening the usage dashboard refresh interval to 10 minutes, clarifying phased model access, and intensifying crackdowns on illicit activity. Z.ai said it would accelerate capacity scaling and offered full refunds for Lite and Pro paid users from January 1 to February 21, 2026. The company added one-click plan restoration for mistaken upgrades and a uniform 15-day subscription extension for all subscribers.
GLM-5 strategy and market positioning
GLM-5 was positioned as a flagship open-source model intended to boost Z.ai's valuation. Released ahead of Lunar New Year, the model emphasizes AI agent capabilities and programming skills, and Z.ai claims it has performance close to Anthropic Claude Opus 4.5. The model scales parameters using sparse attention mechanisms to keep costs relatively low while handling long contexts.
On applications, GLM-5 promotes what Z.ai calls "Agentic Engineering" paired with Z Code tools, enabling users to describe needs while the AI autonomously decomposes tasks and completes development workflows. Demonstrations presented by the company covered general agent assistants, interactive platforms, and gaming, and integration with OpenClaw extends into AI agent products.
Z.ai confirmed GLM-5 compatibility with multiple Chinese domestic chips, including Huawei's Ascend units and Iluvatar (Cambrian Technologies), enabling potential expansion into China's government and enterprise markets. Alongside the GLM-5 launch, Z.ai raised prices by 30% and removed certain discounts, a move that compounded user discontent amid the billing controversy.
In capital markets, Z.ai's valuation on the Hong Kong Stock Exchange surged to HKD323.2 billion (approx. US$41.36 billion), setting a record high. Financial disclosures underscored a gap between market valuation and revenue: CNY312.4 million (approx. US$45.12 million) in revenue for 2024 and CNY191 million in the first half of 2025, figures Z.ai and market observers note sit in stark contrast with the current market value.
Industry context
Observers pointed to intensified competition among Chinese large model developers. Moonshot AI and BytePS have released new models while DeepSeek remained inactive, and the GLM-5 upgrade turmoil was seen as a reminder that technological breakthroughs and commercialization must be matched by operational management and clear user communication for sustainable growth in AI enterprises.
Article translated by Jingyue Hsiao and edited by Jack Wu
