# The Collective Retreat of AI Coding Subscriptions: From Carnival to Tightening in One Quarter ## A Shutdown Notice On April 20, GitHub published a blog post with a restrained title: "Changes to GitHub Copilot Individual Plans." The content was much less restrained: - New signups for Pro, Pro+, and Student plans were paused - Usage limits were tightened through both session limits and weekly limits - Claude Opus was removed from the Pro plan The reason was direct: agentic workflows were consuming far more compute than expected, and "a small number of requests can now cost more than the subscription price itself." Translated: too many people are using it too heavily, and we cannot carry the cost. This is not an isolated case. Overseas vendors have been moving in the same direction. --- ## Overseas: Generous and Restrictive at the Same Time Start with OpenAI Codex. It looks like an outlier. In early April, Sam Altman announced that Codex had reached 3 million weekly active users, reset usage limits for all plans, and promised to reset them again for every additional million users. Community users reported four limit resets within ten days. Plus users currently get 10x usage during the promotion period, up from 5x, while Pro users at $100/month get 2x usage. But the promotion ends on May 31, and nobody knows what happens after that. Windsurf moved differently. In March, it changed from its old Credits system to Quotas. The official explanation was that the new approach was "simpler" and "more industry standard," but the community reaction was intense. Some Reddit users called it an "absolute rugpull." The advantage of the old credit system was transparency: you could see exactly how many credits you spent. The new quota system introduces daily and weekly limits, making real consumption harder to predict. Some users reported that their monthly cost jumped from $30 to $195. Anthropic was not idle either. On April 4, Claude Code announced that third-party tools such as OpenClaw could no longer use subscription quotas. Users would need to pay separately. The official reason was that these tools put "disproportionate pressure" on their systems. The impact was direct: what previously worked under a $20 subscription may now require extra spending, and some estimated bills could jump from $20 to $500. Anthropic later sent some users $100-200 in free credits as compensation. Cursor had already shifted from per-request pricing to usage-based pricing last year, and this year it introduced an Ultra tier at $200/month. The trend is similar to what is happening in China: fixed quotas are gradually giving way to pay-as-you-go usage. --- ## China: The Same Story, Faster I have recently been using Zhipu's GLM Coding Plan Lite as an early user. The experience was fine for the first few months, but recently peak hours have become noticeably slower, with delayed responses and occasional failures. In February, Zhipu announced a price increase of at least 30%. Later it issued an apology and compensated users with 15 days of subscription time, which suggests the experience problem had become widespread. Alibaba Cloud Bailian surprised me even more. In February, it entered the market with a very cheap 7.9 yuan/month Lite plan. At the time everyone was talking about how affordable and useful it looked. Less than two months later, the Lite plan was discontinued: new purchases stopped on March 20, and renewals and upgrades stopped on April 13. New users can only buy Pro at 200 yuan/month. From 7.9 to 200 is a big jump. Other vendors are moving too: - Kimi changed billing from per-use to token-based, and the actual usable quota needs careful calculation - Zhipu added weekly limits to GLM Coding Plan, with peak-hour usage deducted at 3x - Tencent Cloud entered the market in March with a first-month 7.9 yuan promotion, ending April 19 Within one quarter, the market moved from subsidizing user acquisition to collective tightening. The speed was faster than I expected. --- ## Why This Is Happening On the surface, every company is changing prices. Underneath, there is a shared structural problem. Agentic coding eats compute budgets. Traditional code completion is a single request: you type a few characters, the model returns a completion, and token usage is predictable. Agent mode is different. A complex task can include planning, decomposition, multi-step execution, parallel subtasks, verification, and retries. Chained together, token consumption can be tens or even hundreds of times larger than traditional completion. GitHub said this plainly: long-running, parallelized agent sessions exceed the resource limits that the original plan architecture could support. The fixed monthly fee model cannot handle this consumption. Traditional SaaS has near-zero marginal cost for an additional user. AI coding is different. Behind every heavy user is real GPU inference time. Agent workflows cause a single user's usage to explode, and a few requests may cost more than the user's subscription fee. Compute supply is genuinely tight. Zhipu mentioned in its price adjustment announcement that rapid growth in users and calls required increased compute investment. Community users have reported unstable rate limits and frequent rule changes across platforms. High-end GPU inference capacity cannot keep up with user growth. This is not one vendor's problem. --- ## Timeline Put the events together and the trend is clear: | Time | Event | | -------------- | ---------------------------------------------------------------------- | | Mid-2025 | Cursor shifts from per-request pricing to usage-based pricing | | Late 2025 | Zhipu launches GLM Coding Plan | | Feb 2026 | Zhipu raises prices by 30%+ and removes first-purchase discounts | | Feb 2026 | Alibaba Cloud Bailian enters with a 7.9 yuan/month plan | | Mar 2026 | Windsurf changes from credits to quotas, causing backlash | | Mar 2026 | Tencent Cloud launches Coding Plan | | Late Mar 2026 | Alibaba Cloud Bailian Lite stops new purchases | | Early Apr 2026 | Anthropic stops Claude Code subscription support for third-party tools | | Early Apr 2026 | OpenAI Codex repeatedly resets usage limits during the 10x promotion | | Apr 2026 | Alibaba Cloud Bailian Lite stops renewals | | Apr 20, 2026 | GitHub Copilot pauses new signups and tightens limits | The pattern is obvious: use low prices to attract users, then tighten limits, raise prices, and cut plans. --- ## What It Means for Developers As someone who uses AI coding tools daily, I have a few direct impressions. The early dividend is disappearing. Whether it was Zhipu Lite or Alibaba's 7.9 yuan promotion, those prices were unsustainable subsidies. Once user numbers rise and agent usage grows, subsidies inevitably retreat. The price advantage locked in by early subscribers will gradually be diluted by quota limits and worse experience. Choices are narrowing. GitHub paused signups, Alibaba cut Lite, and Zhipu removed first-purchase discounts. For new developers entering the field, the threshold is rising. Globally, heavy users' real monthly spending is moving toward $60-200, or roughly 400-1400 yuan. Heavy API users may even reach $500-2000. Billing is becoming more complex. Windsurf moved from credits to quotas, Cursor uses usage-based pricing, Claude Code cut third-party tool support, and Zhipu added multiplier deductions. Session limits, weekly limits, token billing, and model multipliers are all ways vendors control cost. Users now need to read the rules carefully to know what their money actually buys. There is occasionally good news. Codex is currently in a promotional period, with generous quotas and frequent resets. But a promotion is still a promotion. The strategy after the end of May remains unknown. --- ## What May Happen Next Several directions are worth watching. Fixed monthly pricing will gradually give way to tiered billing. GitHub is already moving this way: basic completion stays cheap, agent mode is billed by usage, and high-end models cost extra. Domestic vendors will likely follow. Open-source models may become attractive again. As subscription costs rise, locally deployed open models such as Qwen and DeepSeek regain an economic advantage, assuming you have enough local compute. The gap between individuals and enterprises will widen. Enterprise customers are willing to pay for stability, and enterprise pricing usually comes with better service quality. Individual developers can only choose within ever-tightening quotas. --- ## Closing Thoughts This tightening of AI coding subscriptions is essentially what happens when technical progress outruns business model evolution. Agent capabilities have improved and user experience has become better, but the compute cost behind each agent call is real. When "more users using it more deeply" changes from good news into operational pressure, tightening becomes inevitable. I currently subscribe to Codex, GLM Lite, and Windsurf, and choose between them depending on task complexity. I do not put all my bets on a single subscription. Everyone needs to decide which services to subscribe to based on their own needs. But either way, the 7.9 yuan era is probably gone. --- ## References **GitHub Copilot** - [GitHub Blog: Changes to GitHub Copilot Individual Plans](https://github.blog/news-insights/company-news/changes-to-github-copilot-individual-plans/) (2026-04-20) - [Sina Finance: GitHub Copilot adjusts individual subscription plans](https://finance.sina.com.cn/tech/digi/2026-04-21/doc-inhvfivk6364481.shtml) (2026-04-21) **OpenAI Codex** - [OpenAI Developers: Codex Pricing](https://developers.openai.com/codex/pricing) - [Business Today: OpenAI Codex celebrates 3 million weekly users, CEO Sam Altman resets usage limits](https://www.businesstoday.in/technology/story/openai-codex-celebrates-3-million-weekly-users-ceo-sam-altman-resets-usage-limits-524717-2026-04-08) (2026-04-08) - [Reddit: Codex rate limits changed from 1 week to 2 weeks](https://www.reddit.com/r/codex/comments/1rqm7ll/codex_rate_limits_changed_from_1_week_to_2_weeks/) - [OpenAI Community: Understanding the New Codex Limit System After the April 9 Update](https://community.openai.com/t/understanding-the-new-codex-limit-system-after-the-april-9-update/1378768) **Windsurf** - [Windsurf Blog: Introducing our new pricing plans](https://windsurf.com/blog/windsurf-pricing-plans) - [Windsurf Docs: Quota-Based Usage](https://docs.windsurf.com/windsurf/accounts/quota) - [Reddit: The new "quota" system is an absolute rugpull](https://www.reddit.com/r/windsurf/comments/1ryalzv/the_new_quota_system_is_an_absolute_rugpull/) **Claude Code / Anthropic** - [TechCrunch: Anthropic says Claude Code subscribers will need to pay extra for OpenClaw support](https://techcrunch.com/2026/04/04/anthropic-says-claude-code-subscribers-will-need-to-pay-extra-for-openclaw-support/) (2026-04-04) - [Mashable: Claude subscribers will have to pay extra to use OpenClaw](https://mashable.com/article/claude-users-will-have-to-pay-extra-to-use-openclaw) - [Morph: The Real Cost of AI Coding in 2026](https://www.morphllm.com/ai-coding-costs) **Cursor** - [Cursor Blog: Updates to Ultra and Pro](https://cursor.com/blog/new-tier) - [Vantage: Cursor Pricing Explained 2026](https://www.vantage.sh/blog/cursor-pricing-explained) - [Gate.com: Cursor Faces Backlash Over Pro Plan Pricing Shift](https://www.gate.com/post/status/19808049) **Domestic vendors** - [Sina Finance: Zhipu announces Coding Plan price increase](https://finance.sina.com.cn/wm/2026-02-12/doc-inhmpkar4093196.shtml) (2026-02-12) - [CLS: Zhipu GLM Coding Plan price adjustment](https://m.cls.cn/detail/2287878) (2026-02-12) - [Sina Finance: Zhipu apologizes for GLM Coding Plan issues and announces compensation](https://finance.sina.com.cn/roll/2026-02-21/doc-inhnqzxa1613988.shtml) (2026-02-21) - [Zhipu AI Docs: Coding Plan overview](https://docs.bigmodel.cn/cn/coding-plan/overview) - [Landian News: Zhipu coding plan off-peak quota multiplier adjustment](https://www.landiannews.com/archives/112185.html) - [Alibaba Cloud notice: Coding Plan Lite stops new purchases](https://www.aliyun.com/notice/118094) (2026-03-20) - [Alibaba Cloud notice: Coding Plan Lite stops renewals and upgrades](https://www.aliyun.com/notice/118175) (2026-04-13) - [ITHome: Alibaba Cloud Bailian Lite plan discontinued](https://www.ithome.com/0/938/299.htm) - [Alibaba Cloud Bailian product monthly report, February 2026](https://developer.aliyun.com/article/1713938) - [ITHome: Tencent Cloud launches new Coding Plan](https://www.ithome.com/0/927/699.htm) (2026-03-10) - [Tencent Cloud Coding Plan official campaign page](https://cloud.tencent.com/act/pro/codingplan) - [Zhihu column: From Coding Plan to Token Plan](https://zhuanlan.zhihu.com/p/2023163733262778630) - [China AI Coding Plan benchmark report, March 2026](https://blog.lightnote.com.cn/china-ai-coding-plan-benchmark/)