一張生成的標題卡寫著 Qwen4-Max 對上 Gemini 4,副標題為「更便宜的代幣,更昂貴的任務」,並有三個標籤分別寫著每項任務成本 $3.27 對 $3.16、指數 v5.0.1 43 對 54,以及每項任務時間 32.1 對 14.9 分鐘。
Guides & Insights

Qwen4-Max 對比 Gemini 4:便宜 40% 的 token,對每項任務卻毫無助益

作者

jinhao.song

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

Qwen4-Max 每百萬輸出 token 收費 4.50 美元。Gemini 4 Hydrogen 是 Google 與之對標定價的模型,收費 7.50 美元。根據 Artificial Analysis 自身的測量,完成一項 Intelligence Index 任務在 Qwen4-Max 上要花 3.27 美元,在 Hydrogen 上要花 3.16 美元——在計價表較貴的一端有 40% 的費率優勢,但一旦計入各模型實際消耗的 token,這項優勢就蒸發成 3% 的劣勢。這就是這項比較的核心,而它在任何費率表上都看不到。

There is a complication before any of the numbers mean anything: G​emini 4 is not a model. It is a line of three separately released checkpoints, and only one of them is priced against Qwen4-Max at all. Getting the row wrong is the most common way this comparison goes bad.

Three G​emini 4s, and which one the question is really about

Google has shipped the G​emini 4 generation three times since 30 September 2026, under element names rather than version numbers. Artificial Analysis scores all three, and they are not close to each other:

• Gemini 4 Argon — 於 2026 年 9 月 30 日宣布,自 10 月 13 日起在付費 API 上正式推出。在 Index v5.0.1 上得分 50,輸入 $2.00、輸出 $10.00,每項 Index 任務 $4.97,每項任務 29.2 分鐘,每秒輸出 66.4 個 token。

• Gemini 4 Carbon — 於 2026 年 10 月 28 日宣布,透過限制存取計畫釋出,而非公開註冊。Index 分數為 60,價格為 2.00 美元與 10.00 美元,每項任務 5.48 美元,耗時 34.2 分鐘,每秒 59.9 個 token。本產品線最高分,也是最難取得的一款。

• Gemini 4 Hydrogen — 於 2026 年 11 月 12 日推出,輸入 $1.50、輸出 $7.50,其卡片上未標示切換日期。Index 分數 54,每項任務 $3.16,每項任務 14.9 分鐘,每秒 172.4 個 token——速度是三者中最快的,足足快上兩倍。

Qwen4-Max, generally available since 16 November 2026, is 43 on the same revision at $1.50 and $4.50 in Singapore, $1.413 and $4.24 in the Global-scope regions. So the price comparison people make — Qwen4-Max against "G​emini 4" — is really Qwen4-Max against Argon, because that is the name the generation launched under and the one most coverage anchors to. The capability comparison, though, points somewhere else entirely, because Hydrogen arrived at exactly Qwen4-Max's input price and beats it by eleven index points.

A screenshot of the Artificial Analysis page for Gemini 4 Hydrogen captured 18 November 2026 with an English interface, showing the Google provider tag, an Intelligence Index score of 54, $1.50 input and $7.50 output per million tokens, $3.16 per task, 304M output tokens, 172 tok/s, 14.9 minutes per task and a 42.9-second time to first answer token.

他們真正勢均力敵的那一列

Hydrogen 的輸入價格是每百萬個 token 1.50 美元。Qwen4-Max 在新加坡的輸入價格也是每百萬個 token 1.50 美元。這並非四捨五入的巧合——而是兩家廠商認定同一個數字,正是接近前沿模型的正確切入點——這讓其他每一列都成為決定購買的關鍵。

• 輸出價格 — Qwen4-Max 4.50 美元 vs Hydrogen 7.50 美元。在計價表較貴的一端省下 40%,這也是 Qwen4-Max 唯一明顯較便宜的地方。

• Index,v5.0.1 — Qwen4-Max 43 對 Hydrogen 54,Argon 為 50,Carbon 為 60。十一分的差距代表的是層級變更,而非偏好。

• 每個索引任務成本 — Qwen4-Max $3.27,相比 Hydrogen $3.16、Argon $4.97、Carbon $5.48。一旦計入每個答案的 token 數量,Qwen4-Max 就不是便宜的選項,儘管它在這組中擁有最優惠的輸出費率。

• 完成套件所需的輸出 token 數——Qwen4-Max 127M、Hydrogen 304M,相較於 65.1M 的中位數。兩者都極為冗長;Hydrogen 更是如此,但每項任務的完成成本仍更低,因為其輸出以較低的混合費率計費,且推理 token 比 Carbon 少。

• 每個索引任務的時間 — Qwen4-Max 32.1 分鐘,對比 Hydrogen 14.9、Argon 29.2、Carbon 34.2。這是表中最大的功能差距,且指向 Google。

• 輸出速度 — Qwen4-Max 每秒 54.5 個 token,相較之下 Hydrogen 為 172.4、Argon 為 66.4。

• Context and output ceiling — 1M tokens for all four, but Qwen4-Max caps non-thinking output at 131,072 and puts thinking plus answer inside a shared 262,144-token budget, while the G​emini 4 pages document a 64K standard output ceiling with long-decode generation extending up to 1M on the Argon and Hydrogen cards.

• Modality — Qwen4-Max takes text, image and video. Argon takes text and image. Hydrogen and Carbon each add speech and video on top of that, which means the one G​emini 4 that matches Qwen4-Max on price is also the one that already covers its modality gap.

• 重量與使用權——四項皆已截止。Carbon 是唯一需要提出申請的。

A generated two-column scoreboard titled Qwen4-Max vs Gemini 4 Hydrogen - the scoreboard, with six rows reading index 43 against 54, cost per task $3.27 against $3.16, input price $1.50 against $1.50, output price $4.50 against $7.50, time per task 32.1 against 14.9 minutes, and output speed 54.5 against 172.4 tokens per second.

在測試框架介入之前,單次請求的成本是多少

以 150,000 個 token 的提示詞搭配 4,000 個 token 的回答,按定價表計價、關閉思考,Qwen4-Max 使用新加坡,Hydrogen 使用標準費率。

在 Qwen4-Max 上,輸入為 $0.225,輸出為 $0.018,合計 $0.243。在 Hydrogen 上,輸入為 $0.225,輸出為 $0.030,合計 $0.255。Qwen4-Max 以五美分、即 5% 勝出。

現在讓模型思考。Artificial Analysis 測得 Qwen4-Max 在每項索引任務中輸出 104,940 個 token,其中 68,914 個是推理,而答案則短得多。以 100,000 個 token 的思考預算重新執行相同的 150,000 個 token 提示:Qwen4-Max 收費 $0.225 加 $0.45,即 $0.675。Hydrogen 在相同思考量下收費 $0.225 加 $0.75,即 $0.975。Qwen4-Max 以 31% 勝出。

所以,單一請求的算術仍然有利於 Qwen4-Max,而追蹤器上的每項任務數字則有利於 Hydrogen。兩者都正確,且無法透過把輸出 token 數乘以輸出費率來調和——Hydrogen 產生了 3.04 億個 token,Qwen4-Max 則是 1.27 億個,前者是後者的 2.4 倍,價格卻是 1.67 倍;按那種算術,每次執行應該貴上大約四倍,而不是略便宜。公布的每項任務數字,是九項評估中輸入、快取輸入、推理與答案 token 的加權混合,而天真乘法與實測結果之間的落差,正是值得懷疑的部分。在有人公布兩次執行的 token 拆分之前,請把這兩個每項任務數字視為無法從費率卡重建的實測比較,並把任何由其中一個推導出另一個的文章,視為披著證據外衣的算術。

還沒有人定價的時鐘

Argon 的方案屬於入門優惠價。2026 年 12 月 1 日起,價格將從 $2.00 / $10.00 調整為 $4.00 / $20.00,快取費率也同步翻倍。距離本頁時間點僅十三天,而這項變動對整體比較格局的影響,更勝過其中任何一項基準差異。

目前,Qwen4-Max 相較於 Argon,在七分的能力落差下,可節省 25% 輸入與 55% 輸出。自 12 月 1 日起,同樣的落差下,節省幅度將變成 62.5% 與 77.5%。在 11 月下旬做基準測試、12 月才簽約的團隊,評估的是與實際取得不同的產品。

Hydrogen, by contrast, carries no cutover date on its card, and Carbon's restricted distribution means its $2.00 / $10.00 is not a public decision at all. So the honest reading of "Qwen4-Max vs G​emini 4" depends on which of three clocks you are standing in front of: the Argon clock, which is running; the Hydrogen card, which is not; and the Qwen4-Max card, which is also not — A​libaba has published one flat tier from zero to the context limit with no long-context surcharge and no promotional end date.

What Google's models do that A​libaba's does not, and the reverse

The G​emini 4 line's practical advantage is not the score. It is that all three checkpoints have been running long enough to have published throughput, and 172 tokens per second on Hydrogen is a different class of product from 54.5. For anything in front of a user — an IDE assistant, a support agent, a document tool with a spinner — that difference is the product.

Qwen4-Max's advantages are narrower and more concrete. A single flat price tier across the whole 1M window is one: Google's long-decode generation above the standard output ceiling is documented as an extension rather than a flat rate. Protocol breadth is another — Chat Completions, Responses, A​nthropic Messages and DashScope against a single vendor wire format, which is the difference between changing a base URL and changing a client. Modality is not, and anyone who tells you video input is a Qwen-only row has not read the Hydrogen card, which lists speech and video alongside text and image.

There is also the regional question. Qwen4-Max is metered in six regions including Frankfurt and US Virginia, at $1.413 / $4.24 outside Singapore, with batch inference available in Beijing only. Google's G​emini 4 pages publish a single global card. For a team with EU data-residency requirements the A​libaba figure to quote is the Frankfurt one, not the Singapore headline, and it is 5.8% lower.

從同一個地方呼叫全部四個

OrcaRouter carries two of the four: Gemini 4 Argon and Gemini 4 Carbon are in the catalogue as google/gemini-4-argon and google/gemini-4-carbon at Google's list prices with 0% markup passed through, so the 1 December Argon increase lands here the same day it lands on Google's card rather than at the next billing cycle. Qwen4-Max is not hosted here — it is available through A​libaba's own Model Studio API and several third-party platforms.

A screenshot of the OrcaRouter model page for google/gemini-4-carbon captured 18 November 2026 with an English interface, showing the Google breadcrumb, the google/gemini-4-carbon identifier, text, image, file, audio and video input, vision, audio, tools, JSON and reasoning, a 1,048,576-token context note, and zero data retention availability.

針對這個特定比較,有用的模式正是定價結構所暗示的那一種:困難案例用 Carbon,一般預設用 Argon,直到其價格變動為止,而影片與長快取前綴的工作則用 Qwen 等級的模型,因為 Gemini 系列在那裡沒有答案。這是在同一條請求路徑上做出三種模型選擇,而這正是路由 DSL 存在的目的——將請求分類,送到該列所說它應歸屬的模型,並讓 自動容錯移轉 在供應商某天某個時段出狀況時接手。單一金鑰和一個 OpenAI 相容端點,讓 A/B 測試便宜到足以在十二月切換之前而不是之後進行。

Read as a family, G​emini 4 is not one opponent for Qwen4-Max but three, and Qwen4-Max loses the score row to all of them while winning the output-price row against all of them. The number that actually decides it is $3.27 against $3.16 — a per-task tie that a rate-card comparison would have called a 40% win — followed by 32.1 minutes against 14.9, which no amount of token pricing fixes. The date to circle is 1 December, when Argon stops being the cheap Google option and the only honest comparison left is against Hydrogen, at which point the eleven-point gap and the 6.5× speed gap are the entire argument.

本文中的比較1

根據本文內容識別 · 基準測試:Artificial Analysis · 每日更新