SWE-2評測：Cognition新模型逼近Fable卻只綁Devin | SWE-2 Review: Cognition's Cheap Coding Model, Devin-Only
By Kit 小克 | AI Tool Observer | 2026-09-15
🇹🇼 SWE-2評測：Cognition新模型逼近Fable卻只綁Devin
Cognition在9月10日推出新一代程式碼代理模型SWE-2，主打跑分逼近Anthropic的Claude Fable 5.1，成本卻砍到64%以下。這是本週AI圈討論度最高的開發工具話題之一——一個以開源基礎模型改造、專攻「便宜又好用」的程式碼AI，能不能取代訂閱制的昂貴模型？這篇評測整理SWE-2的實際跑分、訓練方式，以及最重要的：現在能不能用。
SWE-2是什麼？
SWE-2是Cognition旗下AI軟體工程師Devin的新引擎，以Moonshot AI開源的Kimi K3（2.8兆參數）為基礎模型，再用強化學習後訓練而成。跟以往分開訓練「快速版」和「深度版」不同，SWE-2一次訓練就同時產出medium、high、max三種算力等級，省下重複訓練的成本。
SWE-2跑分表現如何？
在FrontierCode 1.1主測試中，SWE-2拿下50.0%，只小輸Claude Fable 5.1的50.9%，落後GPT-6 Astra的53.3%。其他測試包括DeepSWE 1.1拿下73.0%、Terminal-Bench 2.1達92.8%。真正的賣點是成本：Cognition宣稱同等表現下比Fable 5.1便宜64%，比GPT-6 Astra便宜約四分之三。跟前代SWE-1.7相比，medium版本用少58%的代理步驟就達到更高分數，單次執行成本降81%。
為什麼SWE-2只能在Devin用？
這是SWE-2最大的限制：目前只開放在Devin Desktop和Devin CLI，Devin Web與Fusion正在陸續開放，但Cognition沒有公布獨立API、定價或開源權重。換句話說，想用SWE-2的跑分優勢，就得綁定Devin整套工作流程，沒辦法像用OpenAI或Anthropic的API那樣直接接進自己的CI/CD或IDE外掛。對已經在用Devin的團隊是好消息，對想單獨測試模型能力的開發者來說則是卡關。
SWE-2值得試嗎？
如果你已經是Devin用戶，SWE-2幾乎沒理由不升級——同樣的訂閱費用換到更接近頂級模型的表現。但如果你是想比較各家程式碼模型效能的獨立開發者，目前只能透過Devin的介面間接感受SWE-2，沒辦法拿它的權重或API跟其他模型做對照測試。建議先觀望Devin Web/Fusion全面開放後，再評估是否值得為了SWE-2轉換工作流程。
好不好用，試了才知道。
🇺🇸 SWE-2 Review: Cognition's Cheap Coding Model, Devin-Only
Cognition launched its newest coding agent model, SWE-2, on September 10, claiming benchmark scores close to Anthropic's Claude Fable 5.1 at under 64% of the cost. It's one of the most-discussed developer tool stories in AI circles this week — a coding model built on an open base model, optimized to be cheap rather than chase the top spot. This review covers SWE-2's actual benchmarks, how it was trained, and — most importantly — whether you can actually use it yet.
What Is SWE-2?
SWE-2 is the new engine behind Devin, Cognition's AI software engineer product. It's post-trained from Moonshot AI's open Kimi K3 base model (2.8 trillion parameters) using reinforcement learning. Unlike training separate fast and deep models, Cognition trained all three effort tiers — medium, high, and max — in a single RL run, cutting duplicate training costs.
How Does SWE-2 Perform on Benchmarks?
On FrontierCode 1.1 Main, SWE-2 scored 50.0%, just shy of Claude Fable 5.1's 50.9% and behind GPT-6 Astra's 53.3%. It also hit 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1. The real pitch is cost: Cognition claims SWE-2 runs 64% cheaper than Fable 5.1 at comparable performance, and roughly a quarter of GPT-6 Astra's cost. Compared to the prior SWE-1.7, the medium tier uses 58% fewer agent steps while scoring higher, cutting per-run cost by 81%.
Why Is SWE-2 Locked to Devin?
This is the biggest catch: SWE-2 is only available inside Devin Desktop and Devin CLI right now, with Devin Web and Fusion rolling out gradually. Cognition hasn't published a standalone API, pricing, or open weights. That means you can't plug SWE-2 into your own CI/CD pipeline or IDE extension the way you would with an OpenAI or Anthropic API — you have to adopt the whole Devin workflow to get the cost advantage.
Is SWE-2 Worth Trying?
If you're already a Devin user, there's little reason not to upgrade — same subscription, closer-to-frontier performance. But if you're an independent developer comparing coding models head-to-head, you can only experience SWE-2 indirectly through Devin's interface, with no weights or API to benchmark against competitors. It's worth waiting for the full Devin Web/Fusion rollout before deciding whether it's worth switching workflows.
好不好用，試了才知道 (You won't know if it's good until you try it).
Sources / 資料來源
- Cognition: Introducing SWE-2 — Pushing the Pareto Frontier
- Tech Times: Cognition SWE-2 Beats Frontier Coding AI at 64% Lower Cost
- BenchLM: SWE-2 Benchmarks & Context
常見問題 FAQ
SWE-2是什麼公司做的？
SWE-2由Cognition開發，是該公司AI軟體工程師產品Devin背後的最新程式碼模型，以Moonshot AI的Kimi K3為基礎訓練而成。
SWE-2跑分贏Claude或GPT嗎？
沒有。SWE-2在FrontierCode 1.1測試中略輸Claude Fable 5.1（50.0% vs 50.9%），也落後GPT-6 Astra（53.3%），賣點是成本比對手低六成以上。
SWE-2現在能用嗎？怎麼用？
可以，目前開放在Devin Desktop和Devin CLI使用，Devin Web與Fusion正陸續開放，但沒有獨立API或公開權重。
SWE-2跟SWE-1.7差在哪？
SWE-2的medium版本用少58%的代理步驟就達到更高跑分，單次執行成本比SWE-1.7低81%，同時一次RL訓練就產出三種算力等級。
延伸閱讀 / Related Articles
- Fable 5.1評測：快取讀取降75%，代理工作省45% | Fable 5.1 Review: Cache Reads Cut 75%, Agents Save 45%
- Fugu Ultra v2評測：不練模型改練調度的AI | Fugu Ultra v2 Review: The Model That's Actually a Team
- AI Agent自建軟體評測：McKinsey曝32%企業不買SaaS | Agentic Coding Review: McKinsey Finds 32% Skip Buying SaaS
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
張貼留言