Who is ahead depends on the bill and the job.誰領先,取決於帳單與工作。
Closed models from the US labs, API and open models from Chinese labs, and one European open-weight line. The index column is filled only when this desk checked an Artificial Analysis article. A dash means "go look", not "worse".美國實驗室的封閉模型、中國實驗室的 API 與開放模型,以及一條歐洲開放權重線。指數欄只在本台核對過 Artificial Analysis 文章時才填。破折號表示「去看」,不是「較差」。
Loading the table…正在載入表格…
I need…我需要…
01
Showing every model. Filter by region or weights if the list is too long.顯示全部模型。清單太長時,再用地區或權重篩。
Comparison比較
02 —
Region地區Weights權重
Model模型
Lab實驗室
Region地區
Weights權重
Released發布
Context上下文
In / out per 1M每百萬輸入/輸出
AA indexAA 指數
Prices are list stickers from the checked notes, before tax, caching, and the tokens a model spends thinking. Gemini 4 Argon's $2 / $10 is a launch discount on a $4 / $20 list. Open models show "Local" because the marginal price is electricity.價格是已核對筆記裡的牌價,尚未計稅、快取,以及模型思考所花的 token。Gemini 4 Argon 的 $2 / $10 是 $4 / $20 牌價上的發表折扣。開放模型顯示「本機」,因為邊際成本是電。
Side by side, up to three並排,最多三個
How to read a leaderboard怎麼讀排行榜
03
Effort is part of the score努力等級是分數的一部分
Opus 5.5's 58 is at max effort. The same model at low effort is a different product and a different bill. If a table hides the effort setting, it is not comparable.Opus 5.5 的 58 是 max effort。同一個模型在低努力是另一個產品、另一張帳單。表格若藏起努力設定,就不能比較。
Preference is not correctness偏好不是正確
Arena votes measure what people prefer in a blind chat. A model can win taste and lose a spreadsheet. Use the index for breadth, your own tasks for a decision.競技場投票量的是盲聊裡人喜歡什麼。模型可以贏品味、輸試算表。用指數看廣度,用你自己的任務做決定。
Open weights need a machine開放權重需要一台機器
Kimi K3 on a leaderboard and Kimi K3 on a laptop are not the same offer. The local advisor is the second half of this page.排行榜上的 Kimi K3 與筆電上的 Kimi K3 不是同一個方案。這頁的另一半是本機顧問。
Re-check before you switch換之前再核對
September 2026 shipped a new notable model every few days. The live board is Artificial Analysis. This table is the annotated snapshot.2026 年 9 月每隔幾天就有一個值得注意的新模型。即時榜是 Artificial Analysis。這張表是加上註解的快照。