Gemini 3.7 Flash評測:半價衝刺,主打寫程式與代理任務 | Gemini 3.7 Flash Review: Half-Price Push for Coding Agents
By Kit 小克 | AI Tool Observer | 2026-08-25
🇹🇼 Gemini 3.7 Flash評測:半價衝刺,主打寫程式與代理任務
Google 在 2026 年 8 月 13 日推出 Gemini 3.7 Flash,距離上一版 3.6 Flash 只隔三週,主打「寫程式與代理任務的最強工作機」。這次更新的重點是價格砍半、思考模式重新設計,還有把上下文視窗拉到 100 萬 token,對開發者跟做 AI 代理(agent)的團隊來說,是這個月最值得關注的模型更新。
什麼是 Gemini 3.7 Flash?
Gemini 3.7 Flash 是 Google 目前定位「工作機」的多模態模型,能吃文字、圖片、影片、音訊跟 PDF。Google 官方形容它是「目前最聰明的工作機模型」,特別針對軟體工程、知識工作跟多步驟代理流程做調校,而不是拚原始推理分數。
Gemini 3.7 Flash 多少錢?
這次最大亮點是定價:即日起到 2026 年 12 月 31 日,Gemini 3.7 Flash 每百萬輸入 token 收費 0.75 美元、輸出 token 3.75 美元,是前一代 3.6 Flash 的半價。但這只是早鳥價——2027 年 1 月 1 日起會直接漲回 1.5 美元/7.5 美元,等於價格翻倍。想省成本的團隊,現在到年底是甜蜜期。
Gemini 3.7 Flash 適合做什麼?
官方數據顯示它在 Terminal-bench 2.1 拿下 85.8%、DeepSWE v1.1 有 65.3%、FrontierCode 1.1 Main 則是 43.6%,主打的是 debug 跟一次通過率,而非純推理能力。上下文視窗來到 100 萬 token,單次回覆最長可到 6.6 萬 token,對餵整個 repo 或長文件的代理任務很實用。
三種思考模式怎麼選?
Gemini 3.7 Flash 提供 low/medium/high 三種思考等級,預設是 medium。要注意的是原本的 minimal 檔位已經取消,硬設會直接回錯誤——如果你的程式串接舊參數,記得先改掉,不然直接 fail。
目前 Gemini 3.7 Flash 已上架 Gemini app(Spark 分頁,需要 AI Pro/Ultra 訂閱)、Google Antigravity、AI Studio、Android Studio 以及 Gemini Enterprise,開發者現在就能在 AI Studio 免費試用。
🇺🇸 Gemini 3.7 Flash Review: Half-Price Push for Coding Agents
Gemini 3.7 Flash launched on August 13, 2026 — just three weeks after Gemini 3.6 Flash — and Google is pitching it as its "most intelligent workhorse model yet" for coding and agent work. The headline changes are a temporary price cut, a redesigned thinking-level system, and a 1M-token context window, making it one of the more practical model releases for developers this month.
What Is Gemini 3.7 Flash?
Gemini 3.7 Flash is Google's latest multimodal workhorse model, accepting text, images, video, audio, and PDFs. Google tuned it specifically for software engineering, knowledge work, and multi-step agent workflows rather than chasing raw reasoning benchmarks.
How Much Does Gemini 3.7 Flash Cost?
The biggest hook is pricing: through December 31, 2026, Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens — half the price of Gemini 3.6 Flash. That's an introductory rate, though: on January 1, 2027, pricing doubles to $1.50/$7.50 per million tokens. If you're cost-sensitive, now through year-end is the window to lock in savings.
What Is Gemini 3.7 Flash Good At?
Google's published numbers show 85.8% on Terminal-bench 2.1, 65.3% on DeepSWE v1.1, and 43.6% on FrontierCode 1.1 Main — the model is optimized for debugging and first-pass code accuracy rather than pure reasoning scores. The 1M-token context window (with up to 66K tokens of output) makes it practical for feeding whole repos or long documents into agent pipelines.
Which Thinking Level Should You Pick?
Gemini 3.7 Flash exposes three thinking levels — low, medium, and high — defaulting to medium. One gotcha: the old "minimal" setting has been removed entirely, and calling it now returns an error. If your integration hardcodes that parameter, update it before you hit a wall in production.
The model is already live in the Gemini app (Spark tier, requires AI Pro/Ultra), Google Antigravity, AI Studio, Android Studio, and Gemini Enterprise — developers can try it free in AI Studio today.
Sources / 資料來源
- Google DeepMind: Introducing Gemini 3.7 Flash
- 9to5Google: Gemini 3.7 Flash launches three weeks after last model
- Google AI for Developers: Gemini 3.7 Flash model docs
常見問題 FAQ
Gemini 3.7 Flash 跟 3.6 Flash 差在哪?
主打程式碼與代理任務調校,價格砍半(限時到2026年底),上下文視窗擴大到100萬token,並移除minimal思考模式。
Gemini 3.7 Flash 可以免費用嗎?
可以在 Google AI Studio 免費試用;正式串接 API 則依 token 用量計費。
Gemini 3.7 Flash 半價優惠到什麼時候?
優惠到2026年12月31日,2027年1月1日起價格會直接翻倍(0.75/3.75美元漲到1.5/7.5美元)。
Gemini 3.7 Flash 支援哪些輸入格式?
文字、圖片、影片、音訊、PDF都吃得下,是多模態模型,並非純文字模型。
延伸閱讀 / Related Articles
- Nvidia投資Perplexity評測:估值飆破300億美元 | Nvidia-Perplexity Deal Review: Valuation Jumps to $30B
- OpenAI Astra評測:2000美元解開10道數學懸案 | OpenAI Astra Review: $2K Solves 10 Open Math Problems
- A2A協議評測:Google交出治理權,跟MCP同框了 | A2A Protocol Review: Google Hands Governance to AAIF
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