On April 21st, my Zhipu Pro quarterly subscription expired. I was too busy to renew it. On April 22nd, I discovered that not only could I not renew, but I also had to “scramble for” limited Coding Plan slots every day at 10 AM like a new user. — Missing this renewal turned out to be fortunate, as it revealed the full picture of the plan changes.
Building Trust in Zhipu AI
As an AI developer, I chose Zhipu’s Coding Plan Pro subscription and renewed it for two consecutive quarters. Initially, it was indeed a cost-effective choice — no weekly limits, reasonable pricing, and a competitive domestic LLM option.
Like many developers, I trusted this company. After all, in the “high-burn-rate” AI race, a vendor that launched a “no weekly limit” subscription seemed to have so much sincerity.
Read more: Escaping the “Subscription Trap”: Why Missing My Zhipu AI Renewal Was a Blessing in DisguiseThe Deterioration of Service Quality
However, the user experience gradually became disappointing. Every timeout, every 429 error, was eroding trust:
The “Slow-Motion Hell” During Afternoon Peak Hours: Almost every afternoon, API responses would become extremely slow, often taking several minutes without returning a single token. Worse still, there were instances of complete silence for 15-20 minutes. For developers relying on AI-assisted programming, this crippling latency is fatal.
The “Double Blow” of GLM-5’s Release:
- Zhipu GLM-5 was released on February 11th, 2026 (late night), but Pro users couldn’t use it until around February 13th (a 1-2 day gap)
- By the time Pro users could finally use it, GLM-5 was too slow, and we basically had to revert to using GLM 4.7.
- Customer service’s explanation: “GLM-5 is positioned against Opus, 4.7 against Sonnet, so you should specify different models for load balancing in Claude Code.” — Sounds professional, but the experience was a mess.
The Torment of 429 Errors: According to Zhipu’s official documentation, error codes 1302 (account-level rate limiting) and 1305 (model overload throttling) both belong to the HTTP 429 category. During peak hours, using a graphical interface to call GLM would likely trigger limitations. User communities are filled with complaints like “Avoid Zhipu GLM Coding Plan, daily 429 throttling, unusable!”
The Breaking Point: Policy Changes
On April 21st, my Pro quarterly subscription expired. I was busy that day and missed the renewal.
On April 22nd, when I tried to renew, I discovered a “new world”: Zhipu announced the cancellation of automatic renewal eligibility for the old unlimited weekly Coding Plan. Existing old plan users cannot renew the old unlimited plan after their current cycle expires.
In other words, I had to transition from an “old user” to a “new user,” “scrambling for” limited Coding Plan slots every day at 10 AM.
The most absurd part: Expired API keys don’t consume system resources, yet Zhipu won’t even provide a “grace period renewal” opportunity.
Connecting the Dots: A Pattern Emerges
As I began investigating this “adjustment,” I discovered a pattern: this isn’t an isolated incident, but a systematic issue.
Although satirical content like z-model.com (a satirical site that tracks and archives “The Shelf Life of Zhipu’s Promises” — documenting every broken commitment to developers) has emerged to express community frustration, we should focus on official records and media reports:
- February 12th: Zhipu released the GLM Coding Plan price adjustment letter, with an overall increase of 30% or more, canceling first-purchase discounts
- February 21st: Zhipu released the “GLM Coding Plan Apology Letter,” admitting to “three mistakes” (insufficient rule transparency, too slow GLM-5 grayscale rollout, rough old user upgrade mechanism design), and gradually opened GLM-5 model access permissions
- April 22nd: Zhipu issued a notice in the official user group, announcing the cessation of old plan automatic renewals (less than 30 days after the February 21st “apology”)
This pattern of making attractive promises to attract users, then changing the terms once they’re hooked, is a classic “bait and switch” — and it’s a red flag for any developer considering Zhipu’s services.
I suddenly realized: If I had successfully renewed for a year on April 21st, I would now be “locked in” for a year without any compensation or exit mechanism.
In retrospect, missing that renewal was a lucky escape.
Industry Lessons and Market Impact
I am not alone in this experience. In Kimi’s developer group, a large wave of users from Zhipu’s Coding Plan flooded in. Some were driven to frustration by 429 errors, some were angered by the “daily scramble” system, and others simply lost faith in the company’s integrity.
Why would Zhipu do this?
According to analysis from WeChat public account 「闲散极客」 and multiple media reports:
- Unlimited weekly plans themselves are difficult to sustain long-term, with extremely high difficulty in cost control for vendors.
- Zhipu’s early decision-making errors: Providing large discounts and high promotion rebates in the phase without competitors, with significant deviations in resource carrying capacity assessment.
- Now in a dilemma: Protecting old user rights means bearing losses, while controlling costs means service quality cannot meet standards.
However, this is not a reason for breaking trust.
Industry Observation: The Cost Will Eventually Come Due
As an industry observer, I believe Zhipu’s “operation” this time will pay a heavy price:
- Extremely Low User Loyalty: The domestic market is crowded, competition is fierce, and the threshold for developers to switch AI services is extremely low, with high price sensitivity. Zhipu is losing its most loyal batch of users.
- The Avalanche Effect of Word-of-Mouth: Negative word-of-mouth spreads faster than any marketing campaign. In the era of social media, a single “breach of trust” will be amplified countless times. Satirical websites like z-model.com and warning posts on Linux.do forum are just the beginning.
- Investor Vigilance: The ultimate support for stock price is “user trust” and “business integrity.” When developers start “voting with their feet,” investors will also reevaluate the long-term value of this company.
A Note on Domestic vs. International Versions
Before you consider alternatives, there’s an important distinction to be aware of: Zhipu AI has both a domestic (China) version and an international version of its Coding Plan.
What I used: I was on the domestic version of Zhipu’s Coding Plan. While it’s cheaper than the international version, the trade-off is frequent throttling during peak hours — the very “slow-motion hell” and 429 errors I described earlier.
The international version: Some Chinese self-media marketing accounts claim that you can purchase the international version of Coding Plan using a Chinese bank card (no bank card verification required), with the main difference being the price. This is essentially a hidden price increase by region.
What’s interesting is that the international version seems to have more sufficient computing power — possibly because their servers are overseas and use more advanced GPUs for inference. I don’t know if the international version’s Coding Plan has the same issues as the domestic version.
I welcome international developers to leave comments sharing your experience with the international version of Zhipu’s Coding Plan! Is it more stable? Is it worth the higher price? Your feedback will help all of us make better-informed decisions.
Recommendations for Developers
If you’re currently using Zhipu’s Coding Plan, here’s what I recommend:
- Audit your actual usage: Track downtime and errors over 2 weeks. Calculate the true cost of lost productivity.
- Test alternatives: Try Kimi, ByteDance, Alibaba, or Tencent. Many developers report better stability. Switching takes less than 1 hour.
- Make an informed decision: In this “buyer’s market” of AI, developers have the power. Don’t let loyalty blind you to better options.
Final Thoughts
On April 21st, I missed my renewal. On April 22nd, I discovered this was an “exit signal” given to me.
In retrospect, missing that renewal was a lucky escape.