跳到主要內容

GPT-6.1 Terra評測:電腦操作代理砍半價更快 | GPT-6.1 Terra Review: Cheaper, Faster Computer Agent

By Kit 小克 | AI Tool Observer | 2026-10-31

🇹🇼 GPT-6.1 Terra評測:電腦操作代理砍半價更快

OpenAI 在 10 月 26 日悄悄推出 GPT-6.1 Terra,主打「電腦操作代理」(computer use agent)——讓 AI 直接看畫面、點滑鼠、填表單,處理那些你懶得自己做的瑣事。重點不是跑分創新高,而是價格直接砍半:API 定價每百萬輸入 token 只要 1 美元、輸出 5 美元,是上一代 GPT-5.6 Terra 與 GPT-6.1 Sol 的一半。

GPT-6.1 Terra 定價砍半,任務速度快 3 倍

Terra 的賣點很直白:便宜、快、夠用。根據 OpenAI 公布的數據,Terra 完成任務的平均速度比 GPT-6 Astra 與 GPT-6.1 Sol 快 3 倍以上;在「最大努力」模式下,錯誤率下降 18 個百分點到 4.3%,已經逼近旗艦模型的水準。現在 ChatGPT Plus、Pro、Enterprise、Edu 用戶都能在 Work 與 Codex 模式直接用,開發者也能透過 API 以 gpt-6.1-terra 呼叫。

跑分輸 Claude Opus 5.5,但成本只要九分之一

在 Agents' Last Exam 這項代理任務測試中,GPT-6.1 Terra 拿下 60.0%,不敵 Claude Opus 5.5 的 63.2%。但真正關鍵在成本:Terra 每個任務平均只花 1.01 美元,Opus 5.5 卻要 8.82 美元——差了快 9 倍。如果你的代理要 24 小時跑在背景處理重複性工作,這個差距比那 3 個百分點的跑分更有意義。

  • 適合:長時間在背景執行的瑣事型代理,例如讀取畫面資料、填寫發票、整理表單
  • 不適合:需要最高精度推理或複雜決策的關鍵任務
  • 定價:輸入 1 美元 / 輸出 5 美元(每百萬 token)

老實說:Terra 不是旗艦模型,OpenAI 自己也沒這麼說

這才是整件事最誠實的地方——OpenAI 從沒宣稱 GPT-6.1 Terra 是最強模型,它的定位很清楚:把「AI 代理」從炫技 demo 變成可以長期燒的預算決策。對大多數只是要跑日常瑣事的代理應用來說,便宜又夠快,可能比多那幾個百分點的跑分更實用。

好不好用,試了才知道。


🇺🇸 GPT-6.1 Terra Review: Cheaper, Faster Computer Agent

OpenAI quietly shipped GPT-6.1 Terra on October 26, its latest model built specifically for computer-use agents — AI that looks at your screen, clicks through apps, and fills out the boring forms you don't want to touch. The headline isn't a benchmark record; it's the price cut. API pricing is $1 per million input tokens and $5 per million output tokens, half of what GPT-5.6 Terra and GPT-6.1 Sol cost.

GPT-6.1 Terra: Half the Price, 3x Faster Task Completion

The pitch is simple: cheaper, faster, good enough. OpenAI says Terra finishes tasks more than 3x faster on average than GPT-6 Astra and GPT-6.1 Sol. At maximum effort, its error rate dropped 18 percentage points to 4.3%, closing in on flagship-level reliability. It is live now for ChatGPT Plus, Pro, Enterprise, and Edu users in Work and Codex mode, and developers can call it via the API as gpt-6.1-terra.

Loses to Claude Opus 5.5 on Benchmarks — But Costs 1/9th as Much

On the Agents' Last Exam benchmark, GPT-6.1 Terra scored 60.0%, behind Claude Opus 5.5's 63.2%. But the cost tells a different story: Terra averages $1.01 per task versus Opus 5.5's $8.82 — nearly a 9x gap. If your agent is running 24/7 on repetitive background work, that cost difference matters more than a 3-point benchmark gap.

  • Good fit: background agents doing repetitive chores — reading screens, filling invoices, sorting forms
  • Not a fit: tasks needing top-tier reasoning or high-stakes decisions
  • Pricing: $1 input / $5 output per million tokens

Honest Take: Terra Isn't a Flagship, and OpenAI Isn't Pretending Otherwise

That is the most refreshing part of this release — OpenAI never claimed GPT-6.1 Terra is their best model. It is positioned to turn AI agents from flashy demos into a line item you can actually afford to run all day. For most agent workloads that are just repetitive busywork, cheap and fast enough beats a few extra benchmark points.

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