DeepSeek V4 正式版上線：跑分追平 Gemini，成本砍到 GPT-5.5 的 1/34 | DeepSeek V4 Goes Stable: Ties Gemini, 34x Cheaper Than GPT-5.5

7月 23, 2026

By Kit 小克 | AI Tool Observer | 2026-07-24

🇹🇼 DeepSeek V4 正式版上線：跑分追平 Gemini，成本砍到 GPT-5.5 的 1/34

DeepSeek V4 今天（7/24）結束為期三個月的 preview 階段，正式轉為 stable release。這件事聽起來像小新聞，但對想把 AI 導入正式產線的企業來說，是一個關鍵的「可以放心用了」訊號——preview 版本三不五時改權重、改行為，正式團隊誰敢把客服機器人或程式碼審查工具押在上面？現在這個顧慮沒了。

DeepSeek V4 的跑分與定價：便宜到有點誇張

DeepSeek V4 分成兩個版本：V4 Pro（1.6T 參數，49B 啟用）與 V4 Flash（284B 參數，13B 啟用），兩者都支援 100 萬 token 上下文。定價如下：

V4 Pro：輸入 $0.435／百萬 token，輸出 $0.87／百萬 token
V4 Flash：輸入 $0.14／百萬 token，輸出 $0.28／百萬 token

在 SWE-bench Verified 這個公認最貼近真實工程場景的跑分上，V4 Pro-Max 拿下 80.6%，是目前開源權重模型的最高分，跟 Gemini 3.1 Pro 打平。換算每輸出 token 的成本，DeepSeek V4 比 Claude Opus 4.8 便宜 28.7 倍，比 GPT-5.5 便宜 34.5 倍。

為什麼 stable release 這件事對企業比模型本身更重要

過去半年開源模型的痛點不是能力不夠，而是「今天測好的 prompt，下週權重一更新就跑不動」。DeepSeek V4 正式版代表官方承諾往後行為穩定，不再有 preview 期的隨機震盪，企業終於能把生產環境的自動化流程建立在這上面，而不用整天提心吊膽。

時間點也不是巧合：Moonshot AI 已承諾 Kimi K3 將在 7/27 開源權重，等於整個七月底變成開源模型密集發布週。這波「開源打價格戰」讓 OpenAI、Anthropic 這類閉源大廠的定價壓力越來越大——當你能用約 1/30 的成本拿到接近的工程跑分，切換的誘因會越來越難忽視。

該不該現在就切換？

老實說，跑分是跑分，你的實際場景是另一回事。建議先做三件事：

拿自己團隊真實的 code review、客服對話樣本跑一輪，不要只信官方 benchmark
1M token context 官方數字漂亮，但長文本「大海撈針」表現要自己測，不同模型在長上下文中段容易漏資訊
先在非核心流程（內部工具、草稿產出）上線，觀察一到兩週行為是否真的穩定

好不好用，試了才知道。

🇺🇸 DeepSeek V4 Goes Stable: Ties Gemini, 34x Cheaper Than GPT-5.5

DeepSeek V4 exited its three-month preview phase today (July 24) and moved to a stable release. That sounds like a minor version bump, but for any team trying to run AI in production, it is the signal that actually matters — a preview build that reshuffles weights and behavior every few weeks is not something you bet a customer-facing bot or a code review pipeline on. That objection is now gone.

DeepSeek V4 Benchmarks and Pricing: Almost Absurdly Cheap

DeepSeek V4 ships in two variants: V4 Pro (1.6T parameters, 49B active) and V4 Flash (284B parameters, 13B active), both with a 1M-token context window. Pricing:

V4 Pro: $0.435 / million input tokens, $0.87 / million output tokens
V4 Flash: $0.14 / million input tokens, $0.28 / million output tokens

On SWE-bench Verified — the benchmark closest to real engineering work — V4 Pro-Max scores 80.6%, the highest of any open-weight model and a tie with Gemini 3.1 Pro. Per output token, DeepSeek V4 comes in 28.7x cheaper than Claude Opus 4.8 and 34.5x cheaper than GPT-5.5.

Why Stable Matters More Than the Model Itself

The real pain point with open-weight models over the past six months has not been capability — it has been the prompt that worked last week breaking after a silent weight update. A stable DeepSeek V4 release means the vendor is committing to consistent behavior going forward, no more preview-era volatility. That is what actually lets enterprises put production automation on top of it instead of babysitting it.

The timing is not coincidental either: Moonshot AI has already committed to open-sourcing Kimi K3's weights on July 27, making the last week of July the densest open-weight release window the industry has seen. This price war is squeezing closed-source vendors like OpenAI and Anthropic harder by the week — when you can get comparable engineering scores at roughly 1/30th the cost, the incentive to switch gets harder to ignore.

Should You Switch Right Now?

Benchmarks are benchmarks — your actual workload is a different story. Three things worth doing first:

Run your team's real code review or support conversation samples through it, do not just trust the official numbers
The 1M-token context looks great on paper, but long-context "needle in a haystack" retrieval needs your own testing — most models still drop information in the middle of long inputs
Deploy it first on non-critical workflows (internal tools, draft generation) and watch behavior for a week or two before trusting it with anything customer-facing

You will not know if it is good until you try it — 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