DeepSeek V4 正式上線：百萬字元視窗標配，尖峰時段變貴兩倍 | DeepSeek V4 Goes GA: 1M Context Standard, Peak-Hour Pricing Doubles Cost

7月 26, 2026

By Kit 小克 | AI Tool Observer | 2026-07-26

🇹🇼 DeepSeek V4 正式上線：百萬字元視窗標配，尖峰時段變貴兩倍

DeepSeek V4 在 2026 年 7 月 20 日正式從預覽版轉為全面上線（GA），距離 4 月 24 日的預覽版釋出剛好不到三個月。這次更新最大的看點不是跑分，而是它把「百萬字元上下文視窗」直接變成標準配備，同時也第一個把「尖峰時段漲價」這種電力公司的定價邏輯搬進了 AI API 市場。

規格：兩個版本，價格差六倍

DeepSeek V4 這次分成兩個版本：

V4-Pro：總參數 1.6T，實際啟用 49B（MoE 架構），輸入 $0.435/M、輸出 $0.87/M
V4-Flash：總參數 284B，實際啟用 13B，輸入 $0.14/M、輸出 $0.28/M

兩者預設都是 1M token 上下文視窗、最大輸出 384K token。這代表即使是最便宜的 Flash 版本，處理一整份中大型程式碼庫或長文件都不再需要切塊處理，這對做 RAG 或長文件摘要的工程師是實質的成本與工程複雜度雙重下降。

跑分表現：不是第一，但很划算

跑分上，DeepSeek V4 並沒有超越 GPT-5.5 或 Claude Opus 4.7 這些第一梯隊模型，但在數學推理與問答類的評測上，已經追平上一代的頂尖水準（約等於 GPT-5.4）。換句話說，它賣的不是「最強」，而是「用一半甚至更低的價格拿到接近頂尖的表現」，這個定位跟過去幾代 DeepSeek 一致。

尖峰定價：AI API 市場的新玩法

真正引發討論的是尖峰時段定價機制。DeepSeek 把北京時間上午 9:00–12:00、下午 14:00–18:00 訂為尖峰時段，這段時間的 API 呼叫費用是離峰時段的兩倍。這是主流 AI 廠商第一次把電力公司常見的「尖峰用電加價」邏輯搬進大型語言模型的計費模式。

對開發者來說，這代表：

批次任務、非即時的資料處理（例如夜間跑的分析腳本）可以刻意排到離峰時段，省下近一半費用
面向使用者的即時應用（客服機器人、程式碼助手）如果集中在工作時間使用，成本會被尖峰定價放大
這可能會成為其他廠商跟進的定價範本，值得持續觀察 OpenAI、Anthropic 是否推出類似機制

該不該換？

如果你的應用場景是長文件處理、程式碼庫分析，或是可以排程到離峰時段執行的批次任務，DeepSeek V4 的性價比相當有吸引力。但如果是即時互動、尖峰時段用量集中的產品，實際成本可能沒有官方標價看起來那麼便宜，上線前建議先用真實流量模式試算一週的帳單。

好不好用，試了才知道。

🇺🇸 DeepSeek V4 Goes GA: 1M Context Standard, Peak-Hour Pricing Doubles Cost

DeepSeek V4 moved from preview to general availability on July 20, 2026, just under three months after its April 24 preview launch. The headline is not the benchmark score. It is that a 1-million-token context window is now the standard default, and DeepSeek became the first major AI lab to import utility-style peak-hour pricing into the LLM API market.

Two Variants, Six Times the Price Gap

DeepSeek V4 ships in two flavors:

V4-Pro: 1.6T total parameters, 49B active (MoE), $0.435/M input, $0.87/M output
V4-Flash: 284B total parameters, 13B active, $0.14/M input, $0.28/M output

Both default to a 1M-token context window with 384K max output. That means even the cheaper Flash tier can now ingest a mid-sized codebase or a long document in one pass, no chunking pipeline required. For teams building RAG systems or long-document summarization, that is a real cut in both cost and engineering complexity.

Benchmarks: Not the Leader, But Efficient

DeepSeek V4 does not top GPT-5.5 or Claude Opus 4.7 on the hardest reasoning tasks. But on math and QA benchmarks, it sits roughly on par with the previous generation's frontier, around GPT-5.4 level. It is not selling "the smartest model." It is selling near-frontier performance at a fraction of the price, which has been DeepSeek's positioning since its earlier releases.

Peak-Hour Pricing: A New Playbook for AI APIs

The real talking point is peak-hour pricing. DeepSeek now charges double the base rate during Beijing-time peak windows, 9:00 a.m. to noon and 2:00 p.m. to 6:00 p.m., mirroring how power companies price electricity during high-demand hours. It is the first time a major LLM provider has applied time-of-day surge pricing to API calls at scale.

What this means in practice:

Batch jobs and non-real-time workloads (overnight analysis scripts, bulk data processing) can be scheduled into off-peak windows to cut costs nearly in half
User-facing real-time products (chat support, coding assistants) that see heavy daytime traffic will absorb the peak surcharge automatically
Watch whether OpenAI or Anthropic follow with similar dynamic pricing, this could become an industry template rather than a one-off

Should You Switch?

If your workload is long-document or codebase analysis, or anything that can be scheduled to run off-peak, DeepSeek V4's price-to-performance ratio is genuinely compelling. But for real-time, daytime-heavy products, the actual bill may run higher than the sticker price suggests. Model a week of realistic traffic against the peak schedule before committing.

好不好用，試了才知道。

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

搜尋此網誌

AI小貼士