Cognition融資評測：Devin估值三個月飆破470億美元 | Cognition Funding Review: Devin Hits $47B in 3 Months

By Kit 小克 | AI Tool Observer | 2026-09-02

🇹🇼 Cognition融資評測：Devin估值三個月飆破470億美元

Cognition AI（AI編程代理Devin的開發商）傳出正在完成一輪約10億美元的融資，估值衝上470億美元，這是Bloomberg在2026年9月2日報導的最新進度。距離上一輪融資才過三個月——今年5月Cognition才以260億美元估值募了10億美元——這次估值幾乎翻倍，投資人對這輪的認購需求高達近100億美元。

Cognition估值三個月翻倍，錢從哪裡來？

Cognition的估值成長曲線相當誇張：2025年9月是102億美元，2026年5月漲到260億美元（pre-money 250億），現在傳出要衝到470億美元。市場解讀是，隨著另一家指標性編程代理新創逐漸被視為「已經名花有主」，Cognition成了AI編程代理賽道少數還能獨立投資的大型標的，資金自然湧入。

Devin的營收撐得住這個估值嗎？

Cognition的年化營收（ARR）確實漲得快：2025年5月只有3700萬美元，一年後（2026年5月）衝上4.92億美元，主要靠併購Windsurf（後改名為Devin Desktop，2026年6月正式上線）把客群從個人開發者訂閱擴大到企業合約。如今公司年化營收據報已突破9億美元，目標是2026年底前跨過10億美元大關。不過用470億美元估值除以9億美元營收，倍數落在約52倍，這在AI編程代理這個競爭激烈、客戶容易換工具的領域，算是相當進取的賭注。

Devin真的有那麼神嗎？老實說評測

撇開融資數字，Devin的實測表現一直有爭議。2024年首次亮相時，官方公布的SWE-bench通過率是13.86%，當時就有工程師質疑示範影片裡的案例其實沒有真正解決問題。第三方測試顯示，在範圍明確的bug修復任務上，Devin的成功率可以到78%，但一旦專案涉及50個以上檔案的跨模組修改，失敗率就明顯上升，容易卡在迴圈裡或做出錯誤假設——曾有測試案例是Devin寫的資料庫遷移腳本差點刪掉來源表的資料，如果沒人審查會直接造成生產環境資料遺失。用戶口碑上，Cognition在Trustpilot只有3.0分，遠低於GitHub Copilot的4.5分和Cursor的4.7分。

小克怎麼看

這輪融資反映的其實是市場對「AI工程師取代人類」敘事的信心，而不是Devin目前的實際能力已經到位。如果你的團隊考慮導入Devin這類AI編程代理，務必先從範圍明確、單一檔案的bug修復開始試水溫，別一開始就丟給它跨模組的大重構——那正是它目前最容易翻車的地方。

好不好用，試了才知道。

🇺🇸 Cognition Funding Review: Devin Hits $47B in 3 Months

Cognition AI, the startup behind AI coding agent Devin, is reportedly closing a funding round of about $1 billion at a $47 billion valuation — news Bloomberg broke on September 2, 2026. That's barely three months after its last round, when Cognition raised $1 billion at a $26 billion valuation in May. Investor demand for this round reportedly hit nearly $10 billion.

Cognition's Valuation Doubled in 3 Months — Why?

Cognition's valuation trajectory is steep: $10.2 billion in September 2025, $25 billion pre-money by May 2026, and now reportedly $47 billion. Part of the story is competitive scarcity — with the sector's other marquee independent coding agent startup increasingly seen as spoken for, Cognition has become one of the few large, independent bets left in the AI coding agent space.

Does Devin's Revenue Back Up the Valuation?

Cognition's annualized revenue (ARR) grew fast: from $37 million in May 2025 to $492 million by May 2026, largely driven by its acquisition of Windsurf (rebranded Devin Desktop in June 2026), which pushed the company from individual-developer subscriptions into enterprise contracts. ARR has reportedly since crossed $900 million, with a stated goal of hitting $1 billion by the end of 2026. Still, a $47 billion valuation against roughly $900 million in ARR works out to about a 52x revenue multiple — an aggressive bet in a category where customers can switch coding agents easily.

Is Devin Actually That Good? An Honest Look

Set the funding numbers aside and Devin's track record is mixed. At its 2024 debut, Cognition cited a 13.86% SWE-bench resolve rate, and engineers at the time questioned whether the flagship demo had actually solved the task shown. Third-party testing since then puts Devin's success rate on narrowly scoped bug fixes around 78%, but failure rates climb sharply on tasks touching 50+ files — it can get stuck in loops or work from wrong assumptions. One test run had Devin write a database migration script that would have deleted data from a source table, a mistake that could cause real production data loss without human review. User sentiment reflects the gap: Cognition sits at 3.0/5 on Trustpilot, well below GitHub Copilot's 4.5/5 and Cursor's 4.7/5.

Kit's Take

This round says more about investor conviction in the "AI replaces engineers" narrative than about what Devin can reliably do today. If your team is evaluating Devin or a similar AI coding agent, start with narrowly scoped, single-file bug fixes before handing over cross-module refactors — that's still where these tools are most likely to fall apart.

好不好用，試了才知道 — you only know if it works once you've tried 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

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

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士