一張生成的標題卡寫著「Qwen 4 對 Gemini 4 Argon:比七分更具決定性的 4.3%」,副標題為「Gemini 4 Argon 在其答錯的題目中有 4.3% 會產生幻覺;Qwen4-Max 則為 28.4%」。三列編號列分別寫著「索引分數 42.9 對 50.4,Artificial Analysis v5.0.1」、「每項完成的索引任務成本 3.27 美元對 4.97 美元」以及「幻覺率 28.4% 對 4.3%」。頁尾寫著「分數與每項任務成本依據 Artificial Analysis Intelligence Index v5.0.1;價格依各廠商公布的價目表。」白底搭配藍與青色點綴的扁平化向量風格,OrcaRouter 標誌合成於右下角。
Guides & Insights

Qwen 4 對決 Gemini 4 Argon:那 4.3% 比七個百分點更具決定性

作者

Gideon Frost

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

The scoreboard everyone reaches for in this matchup is the headline index: 42.9 for Qwen4-Max against 50.4 for G​emini 4 Argon, seven and a half points in G​oogle's favour. But the number that should stop a deployment decision sits three columns to the right of the one everybody quotes. G​emini 4 Argon hallucinates on 4.3% of the questions it gets wrong. Qwen4-Max hallucinates on 28.4% of them. That is a factor of 6.6, and for anyone shipping output that a human reads, it is worth more than seven index points.

The two figures come from the same suite, Artificial Analysis v5.0.1, and neither is a vendor claim. G​oogle has not published a hallucination rate for Argon, and A​libaba has not published one for Qwen4-Max. The independent measurement is the only one that exists.

兩款機型打入同一個市場定位

Both are million-token, image-capable, text-out flagships from companies that build their own silicon. Qwen4-Max went metered on 16 November 2026 in six A​libaba Cloud regions; G​emini 4 Argon has been on sale since 30 September 2026 through G​oogle's API. The cards are close enough that price is not the differentiator people assume.

• Input, per million tokens — G​emini 4 Argon $2.00; Qwen4-Max $1.50 in Singapore and $1.413 in the other five regions. G​oogle is 33% more expensive on the Singapore comparison and 41% more against the Global rate.

• 輸出,每百萬 — Argon $10.00;Qwen4-Max $4.50 與 $4.24。這是較大的差距:主要數字上為 2.2×。

• Cached input — Argon $0.10 per million; Qwen4-Max $0.075 on a read, plus published write rates of $1.875 (five minutes) and $3.00 (one hour) that G​oogle does not break out the same way.

• 每項已完成索引任務的成本 — Argon 4.97 美元;Qwen4-Max 3.27 美元。儘管輸出速率低 2.2 倍,Qwen4-Max 每項任務僅便宜 34%,因為兩者帳單的大部分都是輸入。

• 有效推理強度——Argon 是在 high 設定下測得;Qwen4-Max 則是在 xhigh 設定下。兩款模型都提供階梯式設定,且這兩個數值都無法與另一模型的預設值相比。

• 上下文 — Argon 上為 1,000,000 個 token;Qwen4-Max 在廠商標示規格上同樣是這一百萬,而追蹤器記錄為 983,616,因為該處的思考模式輸入設有上限。

• Hosts — Argon is served by G​oogle; Qwen4-Max by A​libaba Cloud. Neither is open-weight.

A generated two-column scoreboard titled 'Qwen4-Max vs Gemini 4 Argon - the scoreboard'. The Qwen4-Max column reads index score 42.9, cost per index task $3.27, cents per index point 7.6, input per million $1.50, output per million $4.50, hallucination rate 28.4%. The Gemini 4 Argon column reads index score 50.4, cost per index task $4.97, cents per index point 9.9, input per million $2.00, output per million $10.00, hallucination rate 4.3%. A footer reads 'Scores and per-task cost per Artificial Analysis Intelligence Index v5.0.1; prices per each vendor's published rate card.' The OrcaRouter mark is composited bottom right.

按任務拆分並不利於較便宜的卡片。

拆解 $4.97 和 $3.27,令人意外的是哪一個在思考上花得更多。Argon 在每個索引任務中輸出 116,368 個輸出 token,其中 68,301 個是推理;Qwen4-Max 輸出 104,940 個,其中 68,914 個是推理。推理量幾乎相同。差異在於答案:Argon 每個任務寫下 48,067 個 token 的最終答案,而 Qwen4-Max 是 36,025 個。Argon 在可見的部分更為冗長,而在不可見的部分則同樣冗長。

That is a cheerful anomaly in G​emini's favour, because it means the 2.2× output rate is buying 33% more answer text. The per-task premium of 52% is therefore smaller than the rate card suggests, and for workloads where output dominates — long-form generation, code the model writes itself — the gap narrows further. For workloads where input dominates, which is most RAG and most agent loops, Qwen4-Max's input rate does the work instead.

說得直白一點:這些是基準測試工具模擬出的成本,不是實際帳單。你的 token 組合會不一樣,而兩個模型之間的比例也會隨之變動。

Qwen4-Max 真正領先的地方

在这场对决中,反对 Qwen4-Max 的理由很充分,但并非一致意见,而且有两行的走向相反,这一点很重要。

• GDPval — Qwen4-Max 為 1,808.69,Argon 則為 1,633.51。在具經濟基礎的專業任務上,Qwen4-Max 領先一個在指數上勝過它 7.5 分的模型達 10.7%。這是整項比較中最明顯的單一逆轉。

• {{1}}FrontierCode 2.0{{/1}} — 26.5% 對 25.7%。差距在一分以內;實際上可說打成平手,而這正是在 Argon 的頭條領先理應預示會出現差距的那套測試上。

• AutomationBench — 72.6% 對 77.5%。Argon 勝出,但在這套更接近辦公室自動化、而非軟體工程的測試套件上,僅領先五個百分點。

• Terminal-Bench 4.0 — 48.5% 對 57.1%。在這裡,Argon 的領先是真的:在 shell 驅動的工作上領先八點五個百分點。

• HLE Diamond — 44.9% 對 56.7%。在硬科學上相差十二個百分點。

• Terminal-Bench Science — 23.3% 對 52.9%。這是排行榜上最大的差距,也最能明確區隔兩個模型的研究能力。

• Omniscience 準確度——27.7 對 46.1,且兩者都須套用本文開頭提到的幻覺注意事項。

整體來看,這個模式是:Qwen4-Max 在專業判斷任務上能站得住腳,但在任何需要持續技術推理的任務上則明顯落敗。這比「落後七個指數點」是更狹窄也更實用的說法,而且它指向不同的部署情境。

A headless-browser screenshot of the Artificial Analysis page for Gemini 4 Argon, captured 17 November 2026 in English, showing Google's mark, the label Proprietary model, Released September 2026, the heading Gemini 4 Argon: Intelligence, Performance & Price Analysis, a rank strip reading Intelligence #22 of 81, Performance #72 of 79, Cost #59 of 80 and Verbosity #57 of 80, a statistics row giving a 45.598-second time to first answer token, 66 output tokens per second, $2.00 input and $10.00 output per million tokens, a 95% cache discount, 7.55 billion input tokens and an 11.5% output-to-input ratio, a $4.97 per-task cost, and the summary paragraph describing the model as above average in intelligence and reasonably priced for its band, notably slow per task and somewhat verbose, with text and image input, text output and a 1M-token context window.

為什麼 4.3% 的幻覺率值得付出代價

把這項差異量化。如果審查者能在三十秒內核查一項說法,而捏造的說法一旦進入正式環境就會造成十分鐘的返工,那麼每千則答案的預期審查成本,幾乎完全取決於捏造所占的比例,而不是錯誤所占的比例。Argon 的 4.3% 意味著在其不正確答案集合中,大約每二十三則就有一則需要返工;Qwen4-Max 的 28.4% 則意味著大約每三點五則就有一則需要返工。

相較之下,{{1}}每項任務 52% 的溢價{{/1}}對任何設有人工審查員的工作流程都會消失,因為在幾乎任何負載費率下,審查員的時間都讓 token 支出相形見絀。對任何沒有審查員的工作流程,它則會朝另一個方向消失,因為沒有審查員把關的捏造答案不是成本,而是事故。

The honest caveat: hallucination rate is measured on a fixed question set and is a property of the measurement as much as the model. G​oogle could change Argon's behaviour with a system-prompt revision tomorrow and this number would not move until the suite is re-run. Treat it as a strong prior, not a guarantee.

路由問題,單方面解答

Gemini 4 Argon is on our catalogue as google/gemini-4-argon at G​oogle's list price, $2.00 in and $10.00 out per million, with 0% markup on top — the same rates G​oogle publishes, passed through rather than repriced. Qwen4-Max is not on our catalogue: a qwen4 identifier returns not-found, its route is A​libaba's Model Studio API and several third-party platforms, and the closest Qwen model we serve is qwen/qwen3.8-max at $2.00 and $6.00.

這種不對稱本身,就是大多數團隊在部署問題上的答案。在幻覺欄位勝出的模型,是你今天就能透過既有端點呼叫的那一個;而在 GDPval 上勝出的模型,則需要你再簽一份供應商合約才能試用。如果你想在正式投入前,用自己的資料驗證 GDPval 的那項說法,路由 DSL 讓你用一次呼叫,就能把固定的評估集送進兩條路徑並排比較輸出——這比任一模型每項任務的價格所暗示的實驗成本都更低。

A headless-browser screenshot of the OrcaRouter model page for Gemini 4 Argon, captured 17 November 2026 in English, showing the identifier google/gemini-4-argon, the byline Google, an input price of $2.00 per million, a zero-data-retention badge, the note that zero data retention is available under agreement, and both the OpenAI-compatible and generateContent endpoint paths.

這七點沒有告訴你的事

有兩件事會改變這項比較,而且兩者都不遠。對一個採用針對拒答調校之系統提示的 Qwen4-Max 部署進行獨立幻覺量測,將能檢驗 28.4% 究竟是模型本身的特性,還是取決於查詢方式——尚無廠商公布過這樣的量測。而在更大樣本上重跑 GDPval,則能告訴我們 Qwen4-Max 在專業任務上 10.7% 的領先是真實還是雜訊,因為該基準的信賴區間很寬,而兩個模型差距小到足以讓排序翻轉。

Until either lands, the split is clean and it is not the one the index suggests. Sustained technical reasoning, science, shell work, or anything a customer reads unreviewed: G​emini 4 Argon, and the 52% premium is the least interesting number in the deal. Grounded professional judgement at volume, with a review step you were going to run anyway: Qwen4-Max is 10.7% better on the suite that measures exactly that, and it is 34% cheaper per task while you find out.