Claude Fable 5.5：一支沒人打得開的發布影片，以及一個你無法查看的 Blender 示範

On 16 October 2026, a post on X announced the end of one era of AI-assisted coding and the beginning of another, credited the whole thing to C​laude Fable 5.5, and attached a link that was described as the video of the launch. The post says Fable 5.5 wrote pure Python inside Blender while working alongside ElevenLabs, and that the poster was providing the launch footage. Everything in that sentence is checkable except the part that matters: there is no C​laude Fable 5.5. A​nthropic has never shipped it, never announced it, and never documented it. The newest Fable model the company actually publishes is Claude Fable 5.1, released 1 September 2026 at $10 per million input tokens and $50 per million output tokens with a 1M-token context window, and the model A​nthropic's own docs recommend as the default for most work is C​laude O​pus 5.5, released 22 September. So this is a what-we-know-so-far piece about a signal, not a launch post about a model — and the single most useful thing in it is the difference between the two.

這篇貼文實際上說了什麼，以及它沒有包含什麼

這段訊號文字是一行到底的長句。它以一個表情符號開頭，宣告一種做法的終結與另一種做法的到來，點名 Fable 5.5，描述了一段 demo——在 Blender 內以純 Python 撰寫的程式碼，並涉及 ElevenLabs——最後則表示發文者已提供這次發布的 X 影片。

那就是全部的證據基礎。沒有模型識別碼，沒有上下文視窗，沒有價格，沒有知識截止點，沒有基準測試，沒有生成結果的螢幕截圖，也沒有供應商頁面。甚至連一個可運作的產物都沒有：貼文中的縮短連結指向一個影片永久連結，而該連結必須在已驗證的 X 工作階段中才能開啟，因此這段影片無法觀看、無法獨立描述，也無法拿來與任何事物查核。一部無法開啟的發表影片，並不是發表的證據。那只是在聲稱證據存在於其他地方。

作為對比，這就是你實際上可以呼叫的 Fable 生成，今天在紙上看起來的樣子：

• C​laude Fable 5.5 — vendor documentation: none. Model identifier: none published. Rate card: none. Context window: undisclosed. Callable today: no.

• Claude Fable 5.1 — 廠商文件：是。模型識別碼：claude-fable-5-1。發佈日期：2026 年 9 月 1 日。上下文視窗：100 萬個 token，最大輸出 128K。價格：每百萬個 token 輸入 $10 / 輸出 $50。知識截止日期：2026 年 6 月。今日可呼叫：是。

• C​laude O​pus 5.5 — the model A​nthropic's own choosing-a-model page points most workloads at. Released: 22 September 2026. Price: $4 input / $20 output per million tokens.

那種不對稱就是整篇文章的重點。這三個名稱中，有一個有你可以閱讀的頁面、可以編列預算的價格，以及可以發送請求的端點。另外兩個也同樣有這些東西。差別在於，第三個名稱——也就是這篇貼文所談的那一個——這些全都沒有。

為什麼「一段示範影片」是最薄弱的證據形式

示範是模型供應商能為一個不存在的模型製造出來的唯一產物。你不需要端點、價目表或文件頁面來錄製螢幕。你需要的是一塊螢幕。一段在編輯器中出現程式碼的精美短片，只能證明有人錄下了一段在編輯器中出現程式碼的短片，至於它是用什麼生成的，什麼也證明不了。

對於正在提出的這項具體主張而言，這件事比平時更為重要，因為該主張異常鮮明。撰寫驅動 Blender 的純 Python——場景建構、網格操作、動畫曲線、渲染設定——是一項確實困難的代理式任務，而這正是它成為良好示範素材的原因，也正因如此值得抱持懷疑。模型可以產生語法正確的 Blender Python，卻因為十幾個平凡無奇的原因而在執行時失敗。示範可以反覆執行，直到某一次嘗試成功為止。示範可以在另一個系統所產出的執行過程上加上旁白。這些都不是對發布影片者的指控；這只是在陳述一段影片能證明什麼、不能證明什麼。

同樣的道理也適用於這項主張中關於 ElevenLabs 的那一半。將程式設計模型與語音平台配對，是種看似合理的整合，而且這種整合在完成的片段裡看起來一模一樣，不論是模型端到端地統籌完成，還是人類手工把兩個服務拼湊起來。這項主張裡沒有失效模式。這就是它的問題所在。

已推出的 Fable 模型保證會出現的唯一地方

A​nthropic is unusually consistent about this, which is what makes the negative evidence so clean. When the company ships a model, four things appear at essentially the same moment: a row in the models overview with a model identifier beside it, a price on that row, a documentation page at a predictable path, and a change in the API release notes. Not one of those four is optional, and not one of them can be satisfied by a post on X.

As of 17 October, A​nthropic's models overview lists exactly four current models: Claude Fable 5.1, C​laude O​pus 5.5, Claude Sonnet 5.5 and C​laude H​aiku 5.5. Fable 5.1 sits at $10 / $50 per million tokens with a 1M-token context window, 128K maximum output, adaptive thinking that is always on, a default effort setting of high, a June 2026 knowledge cutoff, and a retirement date no sooner than 1 September 2027. The page below that table lists the legacy models, and Fable 5.5 is not there either. A documentation request for a Fable 5.5 overview returns a not-found page. There is no newsroom post, and nothing in the October API release notes touches the Fable line at all — those notes cover a Compliance API beta, an OpenTelemetry beta, fast mode going generally available for C​laude O​pus 5.5, and scheduled deployments for Managed Agents on AWS. Real platform work, none of it a model launch.

那道牆上最接近真正缺口的情況，是某個名稱可能正被保留，而非遭到否認。模型識別碼成本低廉且未經公告；一家公司可以保留其中一個，而不會與它已公布過的任何內容矛盾。所以，誠實的解讀不是「Fable 5.5 是假的」。而是「Fable 5.5 在今天是一個背後沒有任何表面的名稱」，而這兩者是不同的陳述。第一種無法被否證，因此對你毫無用處。第二種則確切告訴你該觀察什麼。

這個訊號實際上是什麼的一部分

這並非 Fable 5.5 首次被人目擊，而知道它在序列中位居何處，正是將它當作新聞來讀，還是當作鏈條中的一環來讀的差別。先前的說法包括一個據報的排程時段、一項關於流量正被導向比 Fable 5.1 更新之物的聲稱，以及在任何廠商公告之前就出現在雲端市集上的一筆目錄項目。每一項都以相同結構出現：一個聽來具體的斷言、沒有附上任何實證，以及一則引用第一則而非為其補充內容的後續貼文。

10月16日那篇貼文的形式如出一轍，只是附加了更令人印象深刻的框架。它的貢獻不是關於該模型的新證據，而是關於該模型做了什麼的新主張；這讓它離可驗證性更遠一步，而不是更近一步。關於模型存在的主張，至少還能靠廠商頁面出現來證實。關於某個示範看起來有多好的主張，卻永遠無法靠任何事物證實，包括廠商真的推出該模型也一樣。

把這件事當作新聞來讀，是有實際代價的。如果你基於 Fable 5.5 即將推出，而且比你手邊的版本更便宜或更好的假設來編列預算，那你就是憑著一支你無法開啟的影片做出了規劃決策。如果它沒有出現，你會在正需要它不是真的那一刻才發現。

本週你實際能做什麼

對於這類訊號，有用的回應既不是等待它，也不是忽視它。而是要讓你在上線日終將面對的決定——模型替換——在那天到來時不花任何成本。那是路由問題，不是建模問題。

On OrcaRouter, anthropic/claude-fable-5.1 is callable now at the vendor's own list price with 0% markup, so the $10 / $50 rate card above is what you pay rather than a reseller's version of it. That matters more than usual in a leak cycle: the moment a vendor moves a price or opens a new model, the change is live here the same day, because there is no markup layer sitting between the list price and your invoice. If A​nthropic does ship a Fable 5.5 and prices it differently, you are not renegotiating anything to find out what that costs.

The other half of being ready is not being exposed. A migration you can make in an afternoon is a migration you can make on a rumour; a migration that requires a rewrite is one you have to commit to early and defend. Putting the current Fable generation behind one O​penAI-compatible endpoint, with automatic failover across providers and a routing DSL that lets you shift traffic between models by changing a rule rather than a codebase, means the answer to "should we move to Fable 5.5" becomes a question you can test on your own traffic in an afternoon instead of a project you have to schedule. Put the unproven model on a slice, keep the proven one on the rest, and let your own evals decide. Model fusion is available for the cases where the honest answer is that you want two models' strengths in one call rather than one model's.

None of that is a reason to doubt that A​nthropic will eventually ship something called Fable 5.5. It is a reason not to have a plan that depends on when.

事情發生的那一天，你會怎麼知道

Four things, and you should look for them together rather than one at a time, because any one of them alone has a benign explanation. A row in A​nthropic's models overview with a model identifier printed next to it. A price published by A​nthropic rather than inferred from Fable 5.1's rate card by somebody else's tooling. A documentation path that resolves. A line in the API release notes. When those four line up, Fable 5.5 is real, and you can evaluate it on your own traffic the same afternoon.

Until then, treat the video as a video. It is a claim about a demo, from an account with no stated relationship to A​nthropic, pointing at footage nobody outside an authenticated session can open, about a model with no identifier, no rate card and no page. That is not a reason to ignore it — early signals are worth reading precisely because they are early — but it is every reason not to act on it.

底線

The 16 October post is the clearest statement yet of what the Fable 5.5 rumour actually is: a compelling story about model capability, attached to a video that does not open, about a model that A​nthropic does not document. The Fable model you can call today is Claude Fable 5.1 — 1M context, 128K output, $10 / $50, released 1 September 2026 — and the model A​nthropic recommends by default is C​laude O​pus 5.5 from 22 September. Everything else in the Fable 5.5 conversation, including this post, is a claim. Build so that the day it stops being one is a boring day: same endpoint, same integration, new model name.

這些都不是假設：像這次發布這類版本背後那些有紀錄的 AI 失效與代理事件，全都收錄在OrcaRouter AI Incident Archive，其中每筆條目都附有自己的來源、嚴重程度、信心等級與爭議標記。