跳到主要內容

Replit預設模型評測:棄用Claude,改押OpenAI | Replit Default Model Review: Drops Claude for OpenAI

By Kit 小克 | AI Tool Observer | 2026-10-18

🇹🇼 Replit預設模型評測:棄用Claude,改押OpenAI

Replit 在 TechCrunch Disrupt 上由 CEO Amjad Masad 親口證實:Replit Agent 的預設模型已經從 Anthropic 的 Claude 換成 OpenAI,這是過去一年多以來最大的 AI 編碼工具「換腦」事件之一。Replit 曾是 Anthropic 的前五大客戶,這次轉向對整個 AI Agent 生態系都是一個值得觀察的訊號。

發生了什麼事

Masad 表示,Replit 今年 8 月已經和 OpenAI 建立合作關係,而現在 OpenAI 的模型成為 Replit Agent 的預設選項,取代原本的 Claude。他給出的理由是三個:成本、合作關係、以及效能——Masad 稱內部與公開的跑分都顯示 OpenAI 在「生成完整應用程式」這項任務上表現更好。

但事情沒那麼單純

  • Claude 並未被下架:使用者仍可以手動切換回 Anthropic 的模型,Replit 官方說法是「保持模型中立(model agnostic)」。
  • 說法互相矛盾:有消息人士向媒體表示,Replit 與 Anthropic 的合作關係其實仍然存在,只是預設值變了,雙方對外說法並不完全一致。
  • 跑分未公開完整方法論:Masad 提到的「內部跑分」目前沒有公開詳細測試集與評分標準,外部無法獨立驗證。

對開發者的意義:不是誰更強,是誰更划算

對實際使用 Replit 的開發者來說,這次事件最該關心的不是「OpenAI 打贏 Claude」的敘事,而是平台預設值隨時可能因為商業合作而改變。如果你的 vibe coding 工作流依賴特定模型的行為模式(例如 Claude 在長上下文、程式碼風格上的習慣),預設模型換了,輸出品質也可能跟著變。

實務建議:

  • 定期檢查 Replit Agent 目前實際呼叫的是哪個模型,不要假設它永遠不變。
  • 重要專案建議手動鎖定模型版本,而不是依賴平台預設值。
  • 把「平台廠商的模型採購決策」和「模型本身好不好用」分開看——前者是商業考量,後者才是你該實測的東西。

這次 Replit 轉向 OpenAI,也再次證明 AI 編碼工具背後的模型供應鏈相當脆弱,任何一家平台都可能因為成本或合作關係說變就變。好不好用,試了才知道。


🇺🇸 Replit Default Model Review: Drops Claude for OpenAI

Replit CEO Amjad Masad confirmed on stage at TechCrunch Disrupt that Replit Agent default model has switched from Anthropic Claude to OpenAI -- one of the biggest provider swaps in the AI coding tool space this year. Replit was previously a top-five Anthropic customer, so this move is worth watching as a signal for the broader AI agent ecosystem.

What Actually Happened

Masad said Replit struck a partnership with OpenAI back in August, and now OpenAI models are the default for Replit Agent, replacing Claude. He cited three reasons: cost, partnership terms, and performance -- claiming internal and public benchmarks show OpenAI performing better at generating complete applications.

The Fine Print

  • Claude is not gone: Users can still manually select Anthropic models. Replit official line is that it remains model agnostic.
  • Conflicting accounts: A person familiar with the Replit-Anthropic relationship told reporters the partnership is technically still active -- only the default changed. The two sides public framing does not fully line up.
  • Unverified benchmarks: Masad internal and public benchmarks have not been published with methodology or test sets, so outsiders cannot independently verify the performance claim.

Why It Matters: Not Which Model Wins, But Defaults Can Shift

For developers actually using Replit, the real takeaway is not a verdict on OpenAI versus Claude -- it is that platform defaults can change overnight based on business deals, not just model quality. If your vibe-coding workflow depends on a specific model habits (Claude long-context behavior or coding style, for instance), a silent default switch can quietly change your output.

Practical takeaways:

  • Check which model Replit Agent is actually calling -- do not assume the default stays fixed.
  • For important projects, pin the model version manually instead of trusting the platform default.
  • Separate which model the platform chose to buy from which model is actually better for your use case -- the former is a business decision, the latter is something you need to test yourself.

Replit pivot to OpenAI is another reminder that the model supply chain behind AI coding tools is fragile -- any platform can switch providers overnight over cost or partnership terms. You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code