DeepSeek V4 Pro評測：今日起API漲價逾兩倍 | DeepSeek V4 Pro Review: API Prices Double Today

By Kit 小克 | AI Tool Observer | 2026-08-16

🇹🇼 DeepSeek V4 Pro評測：今日起API漲價逾兩倍

DeepSeek V4 Pro 是這幾天開發者圈討論最熱的關鍵字——原因不是新功能，而是漲價。DeepSeek 官方宣布，從今天（8/16）UTC 16:00 起，V4 Pro 與 V4 Flash 的計費方式從固定費率改成「尖峰／離峰」兩段式定價，等於現有帳單即將全面調漲，幅度從 2 倍到近 8 倍不等，正好是本文發稿的當下生效。

尖峰／離峰怎麼算：離峰不是折扣，是墊高後的一半

目前 V4 Pro 的固定費率是每百萬 token 輸入 $0.435、輸出 $0.87，快取命中輸入只要 $0.003625，這個價格已經維持了近三個月。新制上路後，離峰時段（尖峰以外）變成輸入 $0.66、快取 $0.022、輸出 $1.98；尖峰時段（UTC 01:00–04:00、06:00–10:00，剛好涵蓋亞洲白天到歐洲上午）則是輸入 $1.32、快取 $0.044、輸出 $3.96。換算下來，就算挑離峰時段用，輸入貴了近 2 倍、輸出貴了近 3 倍，快取讀取更是貴了將近 8 倍——離峰完全不是「省錢時段」，只是比尖峰便宜一半而已。

加價之後，DeepSeek V4 Pro 到底值不值

V4 Pro 是 8 月 13 日正式從預覽版轉為 GA（版本代號 V4-Pro-0813），同時補上 low／high／max 三段推理強度可調，這是它真正的產品進步，不是價格。跑分上它的強項很明確：LiveCodeBench 拿下 93.5%，贏過 Gemini 的 91.7% 與 Claude 的 88.8%；Codeforces rating 3206 也排在同代模型前段班。但弱項同樣清楚：HLE（Humanity's Last Exam）只有 37.7%，落後 Claude 的 40.0% 與 GPT-5.4 的 39.8%；SimpleQA-Verified 事實查核能力 57.9%，跟 Gemini 的 75.6% 有明顯落差。換句話說，拿來寫程式、跑競賽題目它很猛，但拿來做需要精準事實根據的任務要小心。就算漲價後，尖峰時段的 V4 Pro 輸出成本仍比 Claude Opus 4.7、GPT-5.5 便宜 6～13 倍，只是那個「便宜 29 倍」的爆炸性價比故事，今天正式收斂了。

開發者現在該做什麼

檢查你的呼叫時段：如果專案跑批次任務，把非即時的工作排到離峰時段（UTC 尖峰以外），可以省下一半費用。
重新試算月費：原本用固定費率估的成本，現在要按尖峰／離峰比例重算，尤其是輸出量大的應用（長文生成、agent 多輪對話）漲幅最有感。
別只看價格選模型：如果任務偏重程式碼與邏輯，V4 Pro 漲價後仍划算；如果需要高事實準確度，這次跑分差距提醒你該搭配 Gemini 或其他模型做查核。

小結

DeepSeek V4 Pro 這次調價是今年以來對開發者荷包最直接的一次衝擊，尤其對已經把 DeepSeek 排進生產環境成本模型的團隊，建議立刻重新試算。好不好用，試了才知道。

🇺🇸 DeepSeek V4 Pro Review: API Prices Double Today

DeepSeek V4 Pro is the phrase every developer is searching this week — not because of a new feature, but because of a price hike. DeepSeek announced that starting today, August 16 at 16:00 UTC, both V4 Pro and V4 Flash move from flat-rate billing to a two-tier peak/off-peak pricing model. Every existing bill is about to jump, by anywhere from roughly 2x to nearly 8x — and the change takes effect right as this article goes live.

Peak vs Off-Peak: Off-Peak Isn't a Discount, It's Half of an Inflated Rate

V4 Pro's current flat rate is $0.435 per million input tokens and $0.87 per million output tokens, with cache-hit input at just $0.003625 — pricing that has held steady for nearly three months. Under the new scheme, off-peak (any hour outside the peak windows) becomes $0.66 input / $0.022 cache-hit / $1.98 output. Peak hours (01:00-04:00 and 06:00-10:00 UTC, covering Asia daytime through European mornings) jump to $1.32 input / $0.044 cache-hit / $3.96 output. Do the math and even off-peak usage costs roughly 2x more on input, nearly 3x more on output, and almost 8x more on cache reads compared to today's flat rate. Off-peak isn't a savings window — it's simply half of a raised ceiling.

Is DeepSeek V4 Pro Still Worth It After the Hike?

V4 Pro officially graduated from preview to general availability on August 13 (checkpoint V4-Pro-0813), also gaining adjustable reasoning effort levels: low, high, and max. That's the real product upgrade here, not the pricing. Its benchmark strengths are clear: 93.5% on LiveCodeBench, beating Gemini's 91.7% and Claude's 88.8%, plus a 3206 Codeforces rating that leads its generation. But the weaknesses are just as clear: 37.7% on HLE (Humanity's Last Exam), trailing Claude's 40.0% and GPT-5.4's 39.8%, and 57.9% on SimpleQA-Verified versus Gemini's 75.6%, a real gap in factual retrieval. In short: strong for coding and competitive-programming-style tasks, riskier for anything requiring precise factual grounding. Even at peak pricing, V4 Pro's output cost still runs 6-13x cheaper than Claude Opus 4.7 or GPT-5.5, it's just that the old "29x cheaper" headline is now history.

What Developers Should Do Now

Audit your call timing: shift batch or non-urgent workloads into off-peak UTC windows to cut costs roughly in half.
Re-run your cost model: any budget built on the old flat rate needs recalculating against the new peak/off-peak split. Output-heavy apps (long-form generation, multi-turn agents) will feel this the most.
Don't pick a model on price alone: for code-and-logic-heavy tasks, V4 Pro remains a good deal even after the hike; for fact-sensitive tasks, today's benchmark gap is a reminder to pair it with Gemini or another model for verification.

Bottom Line

This is the most direct hit to developer wallets from any DeepSeek update this year, and teams that already baked DeepSeek into production cost models should re-run the numbers immediately. 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