Gemini 3.7 Flash評測：編碼跑分狂飆但明年漲價一倍 | Gemini 3.7 Flash Review: Cheap Coding Model, Price Hike Set

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

🇹🇼 Gemini 3.7 Flash評測：編碼跑分狂飆但明年漲價一倍

Gemini 3.7 Flash 是 Google DeepMind 於 2026 年 8 月 13 日推出的中階模型，主打「便宜又能打」：程式碼相關跑分大幅超車前代 Gemini 3.6 Flash，價格卻只要一半。但仔細看定價表會發現一個關鍵：現在的低價只是「早鳥優惠」，2027 年 1 月起價格就會直接翻倍。這篇不吹不黑，直接看數據跟眉角。

跑分真的有進步，尤其是程式碼與長文本

Gemini 3.7 Flash 在幾項關鍵跑分上進步明顯：

FrontierCode 1.1：34.4% → 43.6%
Terminal-bench 2.1：78.0% → 85.8%
AutomationBench：17.0% → 30.4%（幾乎翻倍）
長文本理解（GDM-MRCR）：91.8% → 97.0%

context window 拉到 100 萬 token，輸出上限 6.4 萬 token，對於要餵大量程式碼庫或文件做 agentic 任務的場景很實用。在 Code Arena 網頁開發評測上，Gemini 3.7 Flash 甚至贏過 Claude Sonnet 5 和 GPT-5.6 Terra。

但別被跑分沖昏頭：Agent 任務仍落後

誠實講，Gemini 3.7 Flash 不是全面最強。第三方測試指出，在 DeepSWE、Terminal-bench 完整版、OSWorld 這類需要長時間自主操作的 agentic 評測上，GPT-5.6 Terra 依然領先。換句話說，寫程式碼片段、單次任務它很強；但要放給它自己跑一整套多步驟工作流，還是有落差。另外模型輸出目前僅支援文字，即便輸入端能吃圖片、影片、PDF、音訊，也不支援「minimal」思考層級參數，呼叫時會直接報錯，串接前建議先測試好參數組合。

定價陷阱：早鳥價只到年底

這是最容易被忽略的重點。現在的定價是 Input $0.75／Output $3.75（每百萬 token），但這是「2026 年 12 月 31 日前」的早鳥價；2027 年 1 月 1 日起會漲到 Input $1.50／Output $7.50，直接翻倍。如果你打算把 Gemini 3.7 Flash 用於長期生產環境的成本估算，記得把明年的漲價算進去，不要只看現在的促銷數字做財務規劃。

該不該現在導入？

如果你的場景是程式碼生成、程式碼審查、長文本分析這類短週期任務，Gemini 3.7 Flash 目前是同價位帶 CP 值很高的選擇。但如果你需要的是長時間自主運作的 AI Agent（例如全自動化的多步驟工作流），現階段 GPT-5.6 Terra 在對應評測上仍有優勢，值得先做小規模 A/B 測試再決定要不要全面遷移。知識截止日期也要注意：部分領域到 2026 年 3 月，但有些領域仍停在 2025 年 1 月，用在時效性高的任務前務必先驗證。

好不好用，試了才知道。

🇺🇸 Gemini 3.7 Flash Review: Cheap Coding Model, Price Hike Set

Gemini 3.7 Flash is Google DeepMind's mid-tier model released on August 13, 2026, pitched as "cheap and capable" — coding benchmarks jumped significantly over predecessor Gemini 3.6 Flash while pricing was cut in half. But read the pricing table closely and there's a catch: today's low price is an introductory offer that doubles starting January 2027. Here's the honest breakdown, no hype.

Benchmarks Genuinely Improved, Especially Coding and Long-Context

Gemini 3.7 Flash posts clear gains on several key benchmarks:

FrontierCode 1.1: 34.4% → 43.6%
Terminal-bench 2.1: 78.0% → 85.8%
AutomationBench: 17.0% → 30.4% (nearly doubled)
Long-context (GDM-MRCR): 91.8% → 97.0%

The context window now reaches 1M tokens with a 64K output cap, useful for feeding large codebases or documents into agentic workflows. On the Code Arena web-dev benchmark, Gemini 3.7 Flash even edges out Claude Sonnet 5 and GPT-5.6 Terra.

Don't Get Carried Away — Agentic Tasks Still Lag

To be honest, Gemini 3.7 Flash isn't a clean sweep. Third-party evaluations show that on DeepSWE, full Terminal-bench, and OSWorld — benchmarks requiring extended autonomous operation — GPT-5.6 Terra still leads. In short: it's strong for single-shot code generation, but hand it a full multi-step workflow to run unsupervised and gaps appear. Output is also text-only despite multimodal input support, and the "minimal" thinking-level parameter isn't supported — it throws an error on call, so test your parameter combinations before integrating.

The Pricing Trap: Intro Rate Ends This Year

This is the detail most people miss. Current pricing is $0.75 input / $3.75 output per million tokens — but that's the "through December 31, 2026" introductory rate. Starting January 1, 2027, it jumps to $1.50 input / $7.50 output, a full doubling. If you're budgeting Gemini 3.7 Flash for long-term production use, factor next year's price increase into your cost model rather than planning around today's promo number.

Should You Adopt It Now?

For short-cycle work like code generation, code review, or long-document analysis, Gemini 3.7 Flash is currently a strong value pick in its price tier. But if you need a long-running autonomous agent for fully automated multi-step workflows, GPT-5.6 Terra still holds an edge on the relevant benchmarks — worth a small-scale A/B test before committing to a full migration. Also watch the knowledge cutoff: some domains extend to March 2026, others are stuck at January 2025 — verify before using it for time-sensitive tasks.

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

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