週四登場的 GPT-6.2：傳聞背後的儲備重置，以及 OpenAI 究竟說了什麼

At 06:57 UTC today, O​penAI's Codex product lead posted eleven words: "Thursday is feeling codexy to me. Perhaps also a good time for a banked reset." Four minutes later an account with a large following quote-posted it as a launch notice — "Looks like Thursday is finally the big day: prepare your banked resets: the big ship aka GPT-6.2 is incoming" — and by this evening the claim that GPT-6.2 Astra arrives on 12 November is circulating as though it were reported rather than inferred. It is worth being precise about what the two posts contain, because only one is first-party, and the part that is first-party is a sentence about subscription quota, not about a model. If you are holding a production path open for a successor to GPT-6.1 Sol or GPT-6 Astra this week, the difference matters: one of these sentences is a commitment from the vendor and the other is a reader's guess about what a reset implies.

那兩句話，以及它們之間的四分鐘

The primary artifact is small and dated. Thibault "Tibo" Sottiaux, who runs Codex and ChatGPT product at O​penAI, posted at 06:57:08 UTC on 11 November that Thursday felt "codexy" and that it might be "a good time for a banked reset." The community tracker that has followed his reset habit all year logged it within the hour as a teased signal rather than a verified one — the same category it uses for hints, and a different one from the confirmed coupons it stamps when a reset actually lands on its monitoring accounts.

The secondary artifact is the quote-post at 07:01:51 UTC, which is where the model name enters. "GPT-6.2" does not appear in Tibo's post. It does not appear in O​penAI's model documentation, its API changelog or its newsroom either — a catalogue check this evening returns five identifiers in the GPT-6 family and none of them contains 6.2. The label is an inference layered on top of a quota remark, and that inference is doing all the work in the sentence everyone is quoting.

以下是完整的官方原文，供任何想拿來與彙整內容核對的人參考：「週四對我來說挺有 codexy 的感覺。或許也很適合來一次儲備重置。」沒有模型 ID、沒有價格、沒有上下文長度，除了星期幾之外也沒有任何日期。四分鐘後、一則引用轉發之後，這四項全都到齊了。

這些都不構成否定這則貼文的理由。這反而是把它讀成兩項可信度截然不同的主張的理由：一項是第一方陳述，稱週四可能會有重置；另一項是第三方主張，稱週四會是模型發布。前者可對照一年的行為來查核。後者則可對照一份一直未變的目錄來查核。

所謂的「預存重置」究竟是什麼

大多數 Codex 使用者群體之外的讀者，從來不必學這套詞彙，而相關報導卻假定他們懂。儲備重置是付費 Codex 或 ChatGPT 方案中，對 5 小時與每週用量視窗的一次性刷新，並儲存起來備用。它不是折扣，也不是可購買的加購額度：它存放在帳戶中，由使用者決定何時使用，不能轉讓或兌現，而且會過期——今年發放的券通常以三十天為期限。Tibo 親自發放，通常透過郵寄，社群則追蹤每一份。

追蹤器自身截至今晚的數字，讓這個節奏有了輪廓：

• 迄今全球重設次數 — 61 次，最近一次為 10 月 28 日 06:35 UTC，即十四天前。

• 已累積的重置次數 — 12 次，最新一次於十天前進入監控帳戶。

• 節奏——一年下來大約每 4.3 天一次，這使得十四天的間隔成為近來最長的一段沉寂期。

• 今日貼文後的預測——24 小時內 52%，48 小時內 77%，已標示為低信心度。

從那份清單可以得出兩件事。首先，依照追蹤器自身的算法，重置早就該發生了，所以預測週四會有一次重置並不是大膽的判斷——而是基準率終於追上來了。其次，全域重置與存入式重置之間的區別，正是這篇貼文的全部重點。Tibo 說的是「存入式重置」，那是規模較小、更個人化的工具：每個帳戶一張優惠券，而不是替所有人扳動同一個開關。這個系列產品的發布日往往產生的是全域那種。

為什麼「重置預告」是薄弱的模型訊號，具體來說

The temptation is to treat any Tibo post as a launch tell, because he has used resets as launch garnish before. In September, O​penAI reset usage for all paid ChatGPT subscriptions right after GPT-6 Astra finished rolling out to paid tiers, framed as compensation for users who had burned through quota during the surge. That is the pattern people are matching against.

But look at what the reset tracked in that case: a rollout that had already happened, to a model that was already documented, priced and callable. The reset trailed the release. It did not precede it, and it was global because the pain was global. A banked coupon offered ahead of an unknown date is the instrument O​penAI reaches for when it wants goodwill without committing capacity — cheap, personal, reversible, expiring.

「codexy」還有一種與新模型無關的解讀。Codex 是一個有自己發布節奏的產品面向：介面變更、投入程度控制、上下文預設值、沙箱行為。GPT-6.1 Sol 的預設上下文在 Codex 內從 272K 移到 500K，這是產品變更，不是模型變更。產品消息正是產品負責人會用一句說「codexy」而不是版本號的話來預告的那類事情。

這些都不會讓週四的重置顯得不太可能。它讓從「重置」跳到「GPT-6.2 Astra」成為這條鏈中最缺乏依據的一步——而這正是每一則對該貼文的彙整都採取的那一步。

訊號其餘部分的實際價值是什麼

這則貼文把另外三項說法一併塞進同一段話裡，而它們並非都以同樣的方式站不住腳。逐一替它們評分，是最快看出這則訊息有多少是在報導、多少只是把本週頭條重述一遍的方法。

• "Yesterday Fable 5.2" — corroborated. A​nthropic's newsroom listed "Introducing Claude Fable 5.2 and Claude Mythos 5.2" on 9 November. Real, dated, and correctly described as yesterday.

• "Today another G​emini" — corroborated in substance, though not as a new generation. G​oogle's G​emini API changelog records Gemini 4 Carbon, model ID gemini-4-carbon, going generally available on 10 November.

• "Don't forget Qwen4" — unshipped. A​libaba has tier names in circulation but has published nothing that constitutes a Qwen4 release, and no Qwen4 model ID resolves on a public endpoint we can reach.

• "Thursday Astra" — no supporting artifact. Nothing in O​penAI's documentation, changelog or newsroom names a date.

四者之中，一個是確鑿的事實，一個是某個已經上市之物的 GA（正式發布），一個是只有名稱而沒有產品，還有一個是標題裡的主張。這個比例就是這篇文章的誠實總結——也正是為什麼這篇值得現在寫，而不是等到週五。有趣的問題不是日期對不對，而是讀者該拿這麼弱的訊號怎麼辦。

能讓此事定案的四件關鍵物證，以及它們各自的狀態

對於「這款模型是否真實存在」有一套固定的檢核清單，而且不會隨傳聞改變。截至11月11日晚上，GPT-6.2 Astra 的四項檢核都還是空的。

• 一列目錄 — developers.openai.com 列出 gpt-6-astra、gpt-6-luna、gpt-6.1-sol、gpt-6.1-terra 和 gpt-oss。沒有任何包含 6.2 的字串。維持不變。

• 公開定價——沒有。最近新增的是 11 月 4 日的圖像模型，以及隔天推出、需經審核才能使用的安全模型，每百萬個 token 輸入 $20、輸出 $100，正好是旗艦模型的兩倍，也是帶入任何價格推測時的上限訊號。

系統卡——沒有。這裡本應是 Preparedness 認定與各項評估的所在，也是本應由一份文件、而非由一項法案來指名該模型的地方。

• {{1}}獨立分數——無。Artificial Analysis 沒有針對 6.2 識別碼的頁面，這與沒有可供衡量的內容一致。{{/1}}

The one artifact that does exist is a billing string. Developers debugging Codex in late October found the interface offering "GPT-6 Astra" while the response payload came back stamped gpt-6.2-astra. Metering is generated by O​penAI's own system, so the identifier exists somewhere in a serving path. It proves a string. It does not produce a price, a context window or a benchmark, and the usage splits people published from it ranged from 1.6% to 80% depending on the account — one bill, not a rollout figure.

週四前該準備好什麼，以及哪些事不要做

不對稱性正是可用之處。如果週四帶來一次重置、卻沒有模型，你已經建立的一切都不需要改動。如果週四帶來一個以未公布費率提供的模型，圍繞錯誤數字規劃的遷移就會耗掉一個預算週期。所以做法是把發布當成路由決策，而不是專案，並把接下來兩天花在無論哪種情況都成立的事情上。

首先，要了解你在實際會替換的方案層級上，每項任務的成本。GPT-6.1 Sol 每百萬輸入 token 為 2 美元、每百萬輸出 token 為 10 美元，快取輸入為 0.10 美元，上下文視窗為 1,050,000 個 token；提示超過約 272K 個 token 時，適用每百萬輸入 4 美元、每百萬輸出 15 美元的長上下文費率。GPT-6 Astra 為 10 美元／50 美元，另有 20 美元／75 美元的長上下文層級。每月 200 萬輸入與 50 萬輸出 token 的工作負載，在 GPT-6.1 Sol 上花費 9 美元，在 GPT-6 Astra 上花費 45 美元——這樣的五倍差距，並非單靠新版本號就能自動證明其合理。

Second, keep the fallback wired before you need it. On OrcaRouter, GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol and GPT-6 Luna are all callable behind one key at the provider's list price with no markup added, so a vendor price change is live on our side the same day rather than after a renegotiation. That matters most in a week a model might appear: if a 6.2 tier lands at a published rate, the number you see is O​penAI's number, and you can send a slice of traffic to it while the current tier stays behind as the second model in a failover chain. Trying an unproven release on 10% of requests is a decision you can reverse on Friday. Repointing a production path at it is not.

第三，不要為傳聞定價。每一項建立在「6.2 Astra 會以熟悉費率推出」這個假設上的成本預測，都只是在猜測一個尚未有人宣布的決定；而這個秋天已經出現了一個機型——cyber 層級——其費率是旗艦方案的兩倍。對假想後繼機型而言，保守的規劃數字不是今天的價目表，而是今天的價目表，再加上高級層級可能成為範本的可能性。

星期四會告訴你什麼，以及不會告訴你什麼

If a reset lands on 12 November, it confirms the tracker's base rate and tells you Tibo felt like being generous. It does not confirm a model, because a banked coupon is the instrument O​penAI uses precisely when it is not ready to commit to something larger. If a model lands, the confirmation will not be the tweet either — it will be a catalogue row with a rate attached, and that row appears on ship day even for gated releases, as the cyber model demonstrated on 5 November.

今天這個訊號有用的地方，不在於它的預測能力，而在於它顯示出：一句關於配額的第一方說法，如何在短短四分鐘內就變成一個發布日期，而且那句話從頭到尾根本沒提到任何模型。要留意產品目錄，而不是倒數計時。而如果你週四寧願做點別的事，而不是等待，那麼今天已經存在的方案，只需按牌價打一次 API 呼叫就能到手，背後還有第二個模型，以防這週到頭來其實是另一回事。