This article argues that several Chinese “CodingPlan” subscription offerings have become much less attractive as vendors move toward token-based billing, reduce quotas, or keep high monthly prices while failing to deliver consistent capability.
The Market Shift
The author says Tencent and Alibaba have largely abandoned fixed coding plans in favor of token-based charging. In practice, that means less predictability, weaker value, and a feeling that long-term users are being pushed into a more expensive model.
Among the remaining options, the article focuses on Volcengine Ark, Zhipu, MiniMax, and Kimi. The author tested all four on two tasks: compiling a current API pricing table for major models and writing a 50,000-word xianxia novel with ten chapters.
Kimi
Kimi is described as the most promising of the four at first glance, yet still disappointing in practice. Its pricing-table output reportedly lagged badly on current model versions, and the quota drained so fast that the author had to copy unfinished novel text manually before the session budget ran out.
Volcengine Ark
Volcengine Ark is portrayed as slightly more usable than Kimi, but still mediocre. Its information was also outdated, and although quota consumption was better, the initial connection latency and overall model ability were not considered strong enough for serious coding or agent work.
MiniMax
MiniMax performed worst in the author’s view. One test reportedly failed early due to API errors, leaving no usable output, while the novel task missed the requested chapter length by a large margin. Low quota consumption did not matter because the outputs themselves were not dependable.
Zhipu GLM
Zhipu came out best among the four, especially in retrieving newer model information, but it still suffered from severe speed issues. In one run the novel generation stalled after three chapters, and switching to a different model version improved speed only by sacrificing output quality and length compliance.
Why the Author Turns to GPT-5.5
The article concludes that overseas offerings, especially GPT-5.5 and Claude Opus 4.7, deliver a stronger mix of capability, speed, and usable quota at a lower monthly price. The author also highlights Codex desktop and terminal tools as a major practical advantage on the OpenAI side.
The overall argument is less about nationalism than about value and reliability: when a domestic subscription costs more, changes plan rules frequently, and still performs worse on real tasks, users will move to whichever tool actually works.