跳到主要內容

DeepSeek V4-Pro評測:百萬token正式版,主打AI代理任務 | DeepSeek V4-Pro Review: 1M Tokens, Now GA for AI Agents

By Kit 小克 | AI Tool Observer | 2026-08-25

🇹🇼 DeepSeek V4-Pro評測:百萬token正式版,主打AI代理任務

DeepSeek V4-Pro 已在2026年8月13日正式從預覽版轉為正式版(GA),標示為 DeepSeek-V4-Pro-0813,同步開放於 App、網頁與 API。這次更新主打「代理任務」(agentic tasks)——也就是讓 AI 自己呼叫工具、寫程式、執行多步驟工作流程而不需要人類插手,同時把情境窗口拉到 100 萬 token,是目前開源模型裡數一數二的規格。

DeepSeek V4-Pro是什麼?

DeepSeek V4-Pro 是 DeepSeek 最新的旗艦模型,自今年4月進入預覽測試,8月正式轉為 GA 版本。官方把它定位為「代理優先」模型,強調工具呼叫、程式碼執行與多步驟任務的穩定度,而不只是單純的問答能力。

DeepSeek V4-Pro效能如何?

根據 DeepSeek 官方公布的自測數據,V4-Pro 在 Terminal-Bench 2.1 拿下 87.9 分,DeepSWE 拿下 62.7 分,NL2Repo 拿下 61.5 分,這幾項都是偏向「代理寫程式」的測試項目。要注意的是,這些數字是官方自己跑的,還沒有看到大量第三方獨立覆核,實際表現建議自己拿真實任務測一輪再下結論。

值得注意的功能

  • 百萬 token 情境窗口:可處理長文件、長對話,輸出上限也拉到 384,000 token
  • 思考/非思考雙模式:可依任務難度切換要不要開推理鏈
  • 推理強度分級:low / high / max 三檔可調,權衡速度與品質
  • 原生相容 OpenAI Responses API:並內建 Codex 一鍵接入,對已有 OpenAI 生態的開發者相對友善

DeepSeek V4-Pro怎麼用、划算嗎?

官方 API 定價為每百萬輸入 token 0.435 美元(快取未命中)、每百萬輸出 token 0.87 美元,比起 GPT 或 Claude 系列動輒好幾美元起跳確實便宜不少。但要老實講一句:DeepSeek 在 8 月17日已經宣布尖峰/離峰差別定價,離峰輸出價格漲到約 2.25 倍、尖峰時段漲到 4.5 倍,代表「超便宜」這件事不是永久保證,用之前建議查一下當下即時費率。

小結:適合誰用?

如果你在做需要長情境、多步驟工具呼叫的代理應用,且預算敏感,DeepSeek V4-Pro 是值得放進評測清單的開源選項。但代理能力的實測表現、尖峰時段費率波動,這些都要自己跑過才知道是不是真的划算。

好不好用,試了才知道。

常見問題 FAQ


🇺🇸 DeepSeek V4-Pro Review: 1M Tokens, Now GA for AI Agents

DeepSeek V4-Pro officially left preview and went general availability (GA) on August 13, 2026, shipping as DeepSeek-V4-Pro-0813 across the app, web, and API. The release is built around agentic tasks — letting the model call tools, run code, and complete multi-step workflows without human hand-holding — and pushes context length to 1 million tokens, among the largest of any open model right now.

What is DeepSeek V4-Pro?

DeepSeek V4-Pro is DeepSeek's current flagship model, in preview since April and now fully GA as of August. The company positions it as an agent-first model, optimized for reliable tool calling and multi-step execution rather than just chat quality.

How does DeepSeek V4-Pro perform?

On DeepSeek's own reported benchmarks, V4-Pro scored 87.9 on Terminal-Bench 2.1, 62.7 on DeepSWE, and 61.5 on NL2Repo — all agent-coding-flavored tests. Worth flagging: these are self-reported numbers with limited independent verification so far, so treat them as a starting point, not a verdict, until you run your own workload against it.

Notable features

  • 1M-token context window, with output capped up to 384,000 tokens
  • Thinking / non-thinking mode toggle depending on task difficulty
  • Reasoning-effort levels (low / high / max) to trade speed for quality
  • Native OpenAI Responses API compatibility with one-click Codex setup — friendly if you're already in the OpenAI ecosystem

How much does DeepSeek V4-Pro cost, and is it worth it?

Official API pricing is $0.435 per million input tokens (cache miss) and $0.87 per million output tokens — noticeably cheaper than GPT or Claude tiers that start several dollars higher. But here's the honest part: DeepSeek announced peak/off-peak pricing on August 17, pushing off-peak output to about 2.25x and peak-hour output to 4.5x the earlier rate. "Dirt cheap" isn't a permanent guarantee — check current live rates before you commit a production workload to it.

Bottom line: who should try it?

If you're building agentic apps that need long context and multi-step tool calls on a tight budget, DeepSeek V4-Pro deserves a spot on your evaluation list. But real agent performance and peak-hour pricing swings are things you have to test yourself before trusting the marketing numbers.

好不好用,試了才知道 — you only know if it's good once you've actually tried it.

Sources / 資料來源

常見問題 FAQ

DeepSeek V4-Pro是什麼時候正式推出的?

2026年8月13日從預覽版轉為正式版(GA),版本編號為DeepSeek-V4-Pro-0813。

DeepSeek V4-Pro支援多長的情境窗口?

最長支援100萬token的輸入情境,輸出上限可達384,000 token。

DeepSeek V4-Pro的API價格是多少?

官方定價為每百萬輸入token 0.435美元(快取未命中)、每百萬輸出token 0.87美元,但8月17日起有尖峰/離峰差別定價,尖峰時段輸出費率會漲到約4.5倍。

DeepSeek V4-Pro適合拿來做什麼?

官方主打代理任務,例如工具呼叫、程式碼執行與多步驟自動化工作流程,較適合需要長情境且預算敏感的開發者。

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