智譜就GLM Coding Plan問題致歉
💡Zhipu fixes GLM Coding Plan rollout—refunds for users; GLM-5 now scaling up (coding tool users)
⚡ 30-Second TL;DR
有什麼變化
三大錯誤：規則透明不足、GLM-5灰度慢、升級機制粗糙
為什麼重要
透過退款恢復用戶信任；加速GLM-5存取以競爭編碼AI工具。凸顯中國LLM需求暴增下的擴容挑戰。
下一步行動
Check Zhipu dashboard for GLM Coding Plan refund eligibility if on Pro/Lite.
關鍵要點
- •三大錯誤：規則透明不足、GLM-5灰度慢、升級機制粗糙
- •Max全面開放；Pro高峰限流；Lite節後灰度
- •受影響Pro與Lite用戶可申請退款
- •GLM-5發布後流量超預期，擴容跟不上
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 3 個來源。
🔑 增強重點摘要
- •Zhipu AI issued an apology on 2026-02-22 for GLM Coding Plan issues including insufficient transparency in rules, delayed GLM-5 grayscale rollout, and rough upgrades for old users, amid a traffic surge exceeding capacity after GLM-5 launch[1].
- •GLM Coding Plan subscriptions saw a 30%+ price hike effective February 12, 2026, driven by rising demand for code generation, high platform load, and increased computing costs, with GLM-5 debuting overseas at even higher rates[1].
- •GLM-5 achieved top global open-source performance in early 2026, topping leaderboards, enabling 'Agentic Engineering' for complex tasks, replacing foreign models like Claude Opus 4.5 for many users, and driving Zhipu stock up 32%[2][3].
- •Post-GLM-5, the plan reached 'sold out' status with over 150,000 paying users across 184 countries in four months, offering high cost-effectiveness with token output dozens of times higher than overseas competitors[2].
- •Remedies include full access for Max tier users, peak-limited Pro access, post-holiday Lite grayscale, and refunds for affected Pro/Lite subscribers; existing users retain original pricing despite hikes[1].
📊 競品分析▸ Show
| Feature/Model | Zhipu GLM-5 | Claude Opus 4.5/4.6 | GPT-5.3-Codex |
|---|---|---|---|
| Pricing | Coding Plan 30-60% hike (1/7th Claude cost); high token output | Higher cost (~250/month Max sub) | Not specified; premium tiers |
| Benchmarks | Tops global open-source; agentic engineering, IDE compatibility (Cursor/Claude Code) | Strong in long-range tasks/system engineering | Focus on long-range tasks/codex |
| Users/Demand | 150k+ paying users, 184 countries; replaced foreign APIs for pros | Widely used but pricier | Global leader, but costlier per token |
🛠️ 技術深入
- GLM-5 excels in Agentic Engineering, handling complex system-level tasks like full system creation in 25 minutes, surpassing 'Vibe Coding' era[3].
- Perfect tool compatibility with IDEs like Cursor and Claude Code; no invocation errors, unlike competitors (e.g., MiniMax)[2].
- Optimized for code generation/programming assistance; high token output (dozens of times overseas models at same cost); reaches 90-95% of top foreign model capabilities[1][2].
- Supports long-range tasks and engineering, competing directly with Claude Opus 4.6 and GPT-5.3-Codex[3].
🔮 前景展望AI analysis grounded in cited sources
Zhipu's GLM-5 success signals China's shift from low-price subsidies to premium pricing amid rising costs/demand, stabilizing LLM market rates and validating domestic models' global competitiveness in AI coding tools, potentially accelerating 'programmer replacement' in complex engineering[1][2].
⏳ 時間線
📎 來源 (3)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： 36氪 ↗
每週 AI 簡報
每週一封，可隨時退訂。