一張標題卡寫著「Gemini 4 Carbon vs DeepSeek V4 Pro」,副標題為「37 個指數點,每項任務成本為 4.4 倍」,位於兩張卡片上方:Gemini 4 Carbon 為指數 60、每項任務 $7.76;DeepSeek V4 Pro 為指數 23、每項任務 $1.77。
Guides & Insights

Gemini 4 Carbon 與 DeepSeek V4 Pro:37 個指數點與每項任務 4.4 倍的成本,究竟實際換來什麼?

作者

Rowan Sterling

發佈日期

最新模型 · 20查看全部模型 →
基準測試:Artificial Analysis · 每日更新
返回全部文章

Put G​emini 4 Carbon and DeepSeek V4 Pro on the same scale and the result is not close: 60 against 23 on Artificial Analysis Intelligence Index v5.0.1, a 37-point gap between the fourth-best model on the board and one sitting below the median. Put them on the cost scale and it is not close either, in the other direction — $7.76 per index task against $1.77. The interesting question is not which is better. It is which of those two numbers you are actually paying for in the work you have, because for a large share of production traffic the 37 points go unused while the 4.4x bill arrives every month.

Both models are callable today, which makes this a real comparison rather than a roadmap exercise. G​emini 4 Carbon opened to paid API developers on 10 November 2026 after a twelve-day gated rollout. DeepSeek V4 Pro has been generally available since 13 August 2026 as the 0813 checkpoint, with open weights under MIT.

建立在相反假設之上的兩種模型

規格在任何基準測試之前,就已說明了大部分的行為差異。

• Price, standard rate — G​emini 4 Carbon: $2 in / $10 out per million until 30 November 2026, then $4 / $20. DeepSeek V4 Pro: $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak, straight from D​eepSeek's own rate card.

• Cached input — Carbon $0.10 per million now, $0.20 after 1 December, implicit caching only. D​eepSeek $0.022 off-peak and $0.044 at peak, a 97% discount on the fresh-token rate.

• 上下文與輸出 — 兩者皆具備 1M-token 視窗。Carbon 將同步輸出上限設為 65,536 個 token,並允許在長解碼下最多達 1,048,576 個;V4 Pro 上限為 384,000。

• Input modality — Carbon takes text, image, audio, video and PDF. V4 Pro is text only, and D​eepSeek's own pricing page says vision is not supported on v4-pro.

• 權重 — Carbon 是閉源的。V4 Pro 以 MIT 授權發佈,採用 1.6 兆參數的混合專家架構與 490 億個活躍參數,因此你可以自行託管、微調並在授權條款允許下重新發佈。

• 強度控制 — Carbon 提供低、中、高三種思考等級。V4 Pro 則提供低、高、最高三段推理強度調整。

模態這條界線不是註腳。若你的管線會攝入螢幕截圖、掃描版 PDF 或通話錄音,無論價格如何,V4 Pro 都不能成為堆疊中那一部分的模型,比較就只剩一個候選。若你的流量是程式碼、日誌和文字文件,那麼這個差距就真的是選擇問題。

37 個要點,逐一拆解

那麼大的綜合差距,通常是由少數幾項評測在起主要作用,而獨立評測排行榜對這一對模型所顯示的正是如此。Artificial Analysis 以高努力度測量 Carbon,並以最高努力度測量 V4 Pro 0813;這兩個數字都出自該評測機構,而非廠商。

• 每項 Intelligence Index 任務成本 — Carbon $7.76、V4 Pro $1.77。這個比率已將冗長程度納入考量,因此是兩者之間最公平的單一經濟數字。

• 在此指數上消耗的輸出 token——Carbon 2.05 億、V4 Pro 1.19 億,對比中位數 7,280 萬。兩款模型都很「多話」;Carbon 大約多出 1.7 倍,價格差距有一部分正是由此而來,而非僅來自每 token 費率。

• 速度與延遲 — Carbon 每秒輸出 {{1}}80{{/1}} 個 token,首個回答 token 需時 {{2}}66.85 秒{{/2}};V4 Pro 每秒 {{3}}67{{/3}} 個 token,需時 {{4}}67.71 秒{{/4}}。在這個面向上,兩者幾乎完全相同,而且兩者都慢到若沒有串流與耐心,都不適合放在互動式聊天流程中。

• 努力程度敏感度 — Carbon 在高強度下得分 60,中強度 58,低強度 54。V4 Pro 的非推理配置得分 13,而最高強度下為 23。當你把推理調低時,較便宜的模型損失遠更大,因此「便宜地跑」在兩邊都不是免費的切換。

On the vendor side the numbers are loud and unaudited, and they do not overlap enough to compare directly. D​eepSeek's own card for the 0813 build reports Terminal-Bench 2.1 at 87.9, DeepSWE at 62.7, CyberGym at 83.3, SWE-bench Verified at 80.6 and a Codeforces rating of 3206 — all vendor-reported, none reproduced by a neutral party at the time of writing. G​oogle's card for Carbon reports SEC-bench Pro at 88.2%, ExploitGym at 54.8% and CWE-bench at 71.0%. Note what happens if you try to line those up: CyberGym and ExploitGym are different harnesses, Terminal-Bench 2.1 and Terminal-Bench 4.0 are different versions of a moving target, and only one of the two vendors published a cyber score on a benchmark the other also ran. Any chart setting 83.3 beside 54.8 is comparing a scale to a different scale.

Screenshot of the Artificial Analysis model page for Gemini 4 Carbon (High), showing Intelligence Index v5.0.1 of 60 ranked fourth of 74, $4 per million input and $20 per million output, $7.76 cost per index task, 80 output tokens per second, 66.85 seconds to first answer token, and a 1M-token context window.

沒有人編列預算的計費時鐘

D​eepSeek's price is not one price, and this is the most under-reported difference between the two models. V4 Pro bills at half rate outside a defined peak window: peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else — nights, weekends, holidays — is off-peak.

一週五天、每天二十四小時中有七小時,約占日曆週的 21%。如果你的批次工作是由 UTC 的 cron 排程,這很容易對齊。如果你的流量是具有美國或歐洲日間曲線的即時產品,其中相當比例會落在尖峰時段內,而你實際的混合費率會介於 $0.66 和 $1.32 之間,而不是行銷頁面上的數字。

Carbon 的計時方式正好相反。它在 11 月 30 日之前維持不變,之後則倍增。兩種模型,帳單差異的來源截然不同:一種隨一天中的時段變動,另一種隨日曆日期變動。這兩者都無法從單一的每 token 數字看出來;而拿 $4/$20 對比 $0.66/$1.98 的比較,描述的是一邊的最壞情況與另一邊的最好情況。

坦白說:以 Carbon 的 introductory rate(優惠初期費率)對比 V4 Pro 的 peak rate(高峰費率),Carbon 在輸入端是 1.5 倍,輸出端是 2.5 倍。12 月 1 日之後,對比 off-peak(離峰費率),則是 6 倍與 10 倍。每項任務 4.4 倍的成本數字正好落在這兩者之間,而這才是值得用來規劃的數字,因為它考慮了每個模型實際完成任務需要多少 token。

A self-rendered comparison card headed 'Two billing clocks', contrasting Gemini 4 Carbon's $2/$10 introductory rate running to 30 November 2026 and $4/$20 afterwards with DeepSeek V4 Pro's off-peak $0.66/$1.98 and peak $1.32/$3.96, and noting peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.

差距值得付出代價的地方

與其下定論,不如提出四項站得住腳的分野,並對照這兩個模型實際的形態來檢驗。

Cyber and vulnerability work. This is Carbon's home ground and the reason G​oogle gated it to defenders first. Carbon reads code, images and PDFs in one call, which is what a vulnerability triage pipeline needs, and G​oogle's own reported numbers put it at the top of the cyber boards it chose to publish. V4 Pro has respectable self-reported cyber scores and cannot see a screenshot. For a security team, Carbon at 6x the input price is not obviously the wrong call.

大量分類、擷取與路由。 把這個交給 V4 Pro。37 點綜合差距幾乎無法說明模型是否能標記客服工單,或從發票中擷取欄位;而每百萬輸入 $0.66 的便宜模型,早在品質上落敗之前,就先在算術上勝出。多數團隊在這裡都買過頭了。

長上下文文件處理。兩者的視窗都是 1M 個 token,因此在原始容量上打成平手。真正的差異在於模態與輸出上限:Carbon 可直接吃下掃描版 PDF,並輸出最多 1M 個 token;V4 Pro 則需要事先擷取好的文字,且上限只到 384K。如果你的流程已經完成 OCR,V4 Pro 以十分之一的輸出價格,提供相同的 1M 視窗。

任何你可能必須自行託管的東西。只有 V4 Pro 是選項。MIT 授權的權重意味著沒有供應商能撤回該模型、重新定價,或要求必須透過某項方案才能使用——這並非假設性的擔憂:在這一年,Carbon 的存取條件在十四天內變更了兩次,而其價格將於 12 月 1 日再次變動。

路由的答案,附上誠實的但書

這個決策的成熟做法不是二選一。而是讓 V4 Pro 承擔大部分流量,並將棘手案例升級到前沿模型,這正是我們的 routing DSL 存在的目的:一個負責處理流量的低成本分支、一個在置信度低或驗證失敗時觸發的升級條件,以及同時置於兩者前方的一個 OpenAI 相容端點。

What is on that endpoint today is deepseek/deepseek-v4-pro at D​eepSeek's own $0.66/$1.98, alongside google/gemini-4-argon at $2/$10 — Argon being the G​emini 4 model that carries the same 1M window and a 50 on the same index, at half Carbon's post-December output price. G​emini 4 Carbon is not in our catalogue; it is reached through G​oogle's paid API, and a Carbon branch in a pipeline is a direct integration. Because provider list price is passed through with 0% markup across the 200-plus models on the key, a vendor rate change lands the day it happens, which for Carbon means 1 December.

自動容錯移轉,正是改變平價模型成本盤算的關鍵。團隊之所以超額採購前沿等級的運算資源,是因為害怕在那 5% 真正重要的請求上得到糟糕的答案。有了容錯移轉與升級機制,平價模型就能接手它處理得游刃有餘的那 95%,而當某個請求超出它的能力範圍時,只要透過同一個端點就能触及昂貴的模型——不必為了對沖一段你其實可以改為路由的尾部風險,而在每個 token 上支付前沿等級的費率。

Screenshot of the OrcaRouter model page for DeepSeek V4 Pro, showing off-peak pricing of $0.66 per million input tokens and $1.98 per million output tokens, peak pricing of $1.32 and $3.96, a 1M-token context window, MIT-licensed open weights, and a 1.6-trillion-parameter mixture-of-experts with 49 billion active parameters.

什麼會改變這項建議?

Two developments, both plausible within a quarter. If a neutral evaluator runs Carbon on a cyber benchmark that D​eepSeek also published, the two scorecards become comparable for the first time and the gap stops being an inference. And if D​eepSeek ships the V4.1-Pro that its own migration note pointed at, the cheap side of this comparison gets stronger without the expensive side moving — which would make the escalation split even more attractive than it is now.

在那之前,站得住腳的立場並不光鮮:以 V4 Pro 應付大量需求,只有在多模態輸入或網路能力具關鍵作用時才用 Carbon,而且不要在二十天內就會翻倍的價格兩端簽下長期合約。