Atria Dawn 對決 GPT-5.6 Sol：13 分的指數差距，以及一個視若無睹的研究迴圈

Here is the entire tension of this comparison in one line: Atria Dawn Preview is a brand-new open-weights agentic model from the Shanghai AI Laboratory, released on September 14 on a 744B MoE GLM-5.2 base with a 256K context and no published price, while GPT-5.6 Sol is OpenAI​'s flagship tier of the GPT-5.6​ family — shipped July 9, priced at $4.00/$20.00 per million tokens, and carrying an independent Artificial Analysis Intelligence Index of 61 that has been stable through two price cuts. On paper the gap is enormous: a 13-point independent index delta between a closed frontier flagship and an unproven open preview. The reason the matchup is worth reading anyway is that Atria Dawn Preview was not built to close that gap — it was built to make it irrelevant, by targeting the one thing a raw index score does not measure: whether a model will keep working until it has a verifiable result, without a human holding the thread.

Sourcing labels carry this whole article, so they come first. GPT-5.6 Sol's index score, context, and pricing are independent facts you can check on Artificial Analysis and OpenAI​'s own rate page today. Atria Dawn Preview's benchmark rows are vendor-reported from its model card, unreproduced by any neutral lab, and its international API at api.atria-asi.ai has no published price yet. Where the card pits the two against each other, Atria Dawn Preview reports leads on discovery and tool-use suites (BrowseComp 92.5 vs 90.8, BFCL v4 77.0) while GPT-5.6 Sol's coding rows sit far ahead (SWE-bench Pro 59.6 vs a reported 74.7-equivalent on the same grid, Terminal-Bench 2.1 78.3 vs 90.2). None of the Atria numbers is independently verified; treat them as the lab's pitch, not a scorecard.

兩個只共有「模型」這個詞的物件。

GPT-5.6 Sol is OpenAI​'s flagship reasoning tier: a 1.05M-token context window, up to 128K output, text/image/file input, an AA Intelligence Index of 61, and a $4/$20 promotional rate that OpenAI​ cut from $5/$30 on August 24 and has committed to keep at least through November 21 — with a 272K-token cliff above which the entire request bills at $8/$30. It is the model behind Codex, behind Work, and behind the default ChatGPT experience; it is the most widely deployed reasoning model in production right now. It is also proprietary: you rent it, you never own it.

Atria Dawn Preview 是相反的經濟產物。它採用 MIT 授權，提供可下載的權重（BF16 分為 353 個分片，FP8 分為 177 個），純文字、具備 256K 視窗，可透過 SGLang 與 vLLM 部署，並在國際上託管於 api.atria-asi.ai，使用相同的模型 ID。它的賣點不是原始推理能力——而是研究迴圈：分析、設計、使用工具、撰寫並執行程式碼、讀取結果、復原、迭代。該實驗室的四大支柱——探索、創造、交付、資安——全都歸結於這個迴圈。你無法下載 GPT-5.6 Sol；你可以下載 Atria Dawn Preview，並在你擁有的硬體上運行，而這個差異正是該實驗室押下的全部賭注。

• 價格 — Atria Dawn Preview 未發布（國際 API）對比 GPT-5.6 Sol 每 100 萬 $4.00/$20.00（超過 272K 輸入為 $8/$30）

• 上下文 — Atria 256K 純文字 vs GPT-5.6 Sol 1.05M，128K 輸出

• 獨立紀錄 — Atria 對比 AA Index 61 則無，歷經兩次重新定價仍保持穩定

• 開放性 — Atria 採 MIT 授權權重、可自行架設，對比 GPT-5.6 Sol 為專有、僅提供 API

指標落差實際衡量的是什麼——以及它漏掉了什麼

13 分的指數差距確實存在，你不應該輕描淡寫地帶過。如果你的工作負載是困難的推理問題——一個證明、一個棘手的程式錯誤、一份艱澀的法律文件——GPT-5.6 Sol 平均而言會勝過 Atria Dawn Preview，而它的溢價也是據此定價的。但這個指數主要是由簡短、單次的推理任務所構成，而這正是研究迴圈代理試圖改變規則的領域。Atria Dawn Preview 所回報的強項不是簡短答案；而是 BrowseComp、DeepSearchQA 和 AutomationBench——這些是長程套件，模型可以在多輪中閱讀、行動並自我修正。廠商回報的強弱分布（在探索／工具使用上勝出，在編碼上落後）是為迭代工作、而非為該指數調校的模型之特徵。

For a team that already runs GPT-5.6 Sol for multi-step agentic coding, the honest reading is: Atria Dawn Preview is not a replacement, because its coding rows are exactly where it loses. For a team whose pain is the loop itself — the model stopping to ask, losing the thread, never finishing the experiment — the newcomer is aimed at a real gap, and its open weights mean you can find out for yourself whether the loop holds without paying OpenAI​'s rate for the privilege. Through a routing layer such as OrcaRouter, both models sit behind one key — GPT-5.6 Sol at pass-through list price with automatic failover across providers, and, the moment the Shanghai lab publishes a rate for Atria, the same key routes to it too, so a head-to-head eval costs setup time only.

裁決

別把{{1}}指標差距{{/1}}當成事情的全貌，也別把{{2}}廠商網格{{/2}}當成證明——其中一欄經過獨立驗證，另一欄則是實驗室的自述。如果你今天在 GPT-5.6 Sol 上發布軟體，這裡沒有任何內容支持你切換；13 點差距與 1.05M 視窗對你的工作負載而言已經勝出。如果你從事研究或長時程自動化，而失敗情況是模型放棄，Atria Dawn Preview 的開放權重與迴圈優先調校則是值得進行受控評估的合理實驗——在你自己的硬體上、置於容錯移轉之後，且不要將正式生產路徑押在一個尚未經任何中立人士評分的預覽版上。