Fable 5.1密碼評測：44分鐘解370年懸案，隔天被打臉 | Fable 5.1 Cipher Review: 44-Min Solve, Disputed Next Day
By Kit 小克 | AI Tool Observer | 2026-09-18
🇹🇼 Fable 5.1密碼評測：44分鐘解370年懸案，隔天被打臉
Anthropic的Claude Fable 5.1上週在AI圈掀起一波討論：評測機構Vals AI在8月31日發布報告，宣稱Fable 5.1只花44分鐘、17.6萬個token，在完全沒有人類介入的情況下，破解了塵封370年的Cyphral Distich密碼——藏在蘇格蘭作家Thomas Urquhart 1653年著作《Logopandecteision》書末、自1899年就被學界列為未解之謎的64位數字密碼。故事聽起來很燃，但隔天就被另一組研究者踢爆：原始古書裡，可能根本沒有這段密碼。
Fable 5.1是怎麼「破解」密碼的？
Vals AI給Fable 5.1一個開放式任務，只提示它挑一個「可驗證但沒難到不可能」的密碼，並建議它參考自己過去解數學題的經驗，沒有給任何額外提示或步驟。
Fable 5.1最終找到的解法是：64個數字其實是索引，分別對應書中32段「Proquiritations」（誓詞）裡的字詞，取每個被指到的字的第一個字母重組，就拼出一句完整的英文句子：「O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND」。整個搜尋、驗證、排除錯誤假設的過程，沒有人類介入。
隔天被打臉，是怎麼回事？
Vals AI報告發布隔天（9月1日），研究團隊Reticuli Labs就發表技術反駁，指出他們比對大英圖書館館藏的1653年原版後，書末根本沒有這段64位數字密碼——只有印刷裝飾、題辭、「FINIS」字樣跟勘誤表。
這個說法一度讓「Fable 5.1破解密碼」的敘事整個翻盤：如果密碼根本不存在，所謂「44分鐘解謎」就只是模型自己編出一套自圓其說的規則，剛好套用在隨機數字上湊出一句通順的句子。
所以密碼到底破了沒？
後續追查發現，事情比兩邊說法都複雜。蘇格蘭國家圖書館收藏的另一本1653年版本（編號H.32.a.39）確實收錄了這段Cyphral Distich，只是大英圖書館那本剛好沒有——不同印次、不同館藏本內容有落差，在17世紀印刷品裡並不罕見。也就是說，密碼「存在」，但Vals AI最初的報告沒先查證古書的版本差異，就急著發布「AI破解370年密碼」的結論。
另外，Fable 5.1同時嘗試破解一個相關的、篇幅更大的密碼，結果只解開大半，還有9個字母因古書掃描品質或轉錄誤差而無法確定——這是報告裡容易被標題蓋過的細節。
這件事教我們的是什麼？
- AI公司或第三方評測機構發的「破紀錄」跑分新聞稿，標題往往比實際過程更聳動
- 就算模型真的展現出強大的組合搜尋與驗證能力，結論仍要回頭核對一手史料或原始資料才算數
- 對開發者來說，Fable 5.1展現的「長時間自主搜尋、不斷排除假設」能力，比「破解密碼」本身更值得關注——這才是Agent類任務真正吃重的能力
好不好用，試了才知道。
🇺🇸 Fable 5.1 Cipher Review: 44-Min Solve, Disputed Next Day
Anthropic's Claude Fable 5.1 became the AI world's biggest talking point last week after benchmark outfit Vals AI published a report claiming the model cracked the 370-year-old Cyphral Distich cipher — a 64-number puzzle tucked into the back of Scottish writer Thomas Urquhart's 1653 book Logopandecteision, unsolved since scholars flagged it in 1899 — in just 44 minutes and 176,000 tokens, with zero human intervention. Great headline. The problem: a rival research group disputed the very next day that the cipher even exists in the original book.
How Did Fable 5.1 "Solve" the Cipher?
Vals AI gave Fable 5.1 an open-ended prompt, nudging it toward a cipher that was verifiable but not impossibly hard, and suggesting it draw on its own past math work — no other hints or steps were provided.
Fable 5.1's answer: the 64 numbers were indices into the book's 32 "Proquiritations" (oaths) — each number pointed to a word, and the first letter of that word spelled out: "O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND." The whole search, verification, and elimination process ran without a human in the loop.
Why Was It Disputed the Next Day?
On September 1 — the day after Vals AI's report — a group called Reticuli Labs published a technical rebuttal, arguing that after checking a British Library copy of the 1653 original, the book simply doesn't contain this cipher at all; it ends with a printer's ornament, an epigraph, "FINIS," and an errata sheet.
For a moment, that flipped the entire narrative: if the cipher doesn't exist, "solved in 44 minutes" just meant the model invented a self-consistent decoding rule that happened to land on a plausible-sounding sentence from random numbers.
So Was the Cipher Actually Real?
The full picture turned out messier than either side's headline. A separate 1653 copy held by the National Library of Scotland (H.32.a.39) does contain the Cyphral Distich — the British Library's copy simply doesn't. Print-run variation between surviving copies of 17th-century books isn't unusual. So the cipher is real, but Vals AI's original report never checked cross-copy variation before publishing an "AI cracks 370-year cipher" headline.
One detail that got buried under the headline: Fable 5.1 also attempted a larger, related cipher and only partially solved it — nine letters remain uncertain, likely due to transcription errors in the historical scans.
What This Actually Teaches Us
- Benchmark press releases from AI labs and third-party evaluators are often more sensational than the underlying process
- Even genuine search-and-verify capability from a model doesn't make a claim true until someone checks the primary sources
- For developers, the more useful signal here isn't "cracked a cipher" — it's that Fable 5.1 ran a long, autonomous, self-correcting search loop, exactly the skill agentic coding and research tasks actually need
Good tools prove themselves in use, not in a press release.
Sources / 資料來源
- Vals AI: Fable 5.1 Solves the Cyphral Distich
- Schneier on Security: Claude Fable Solves a Historical Cipher
- Reticuli Labs: Distich Refutation Findings (GitHub)
常見問題 FAQ
Fable 5.1真的破解了370年的密碼嗎？
密碼本身確實存在（收藏於蘇格蘭國家圖書館的1653年版本中），Fable 5.1提出的解法也能對應驗證，但最初的評測報告沒查證版本差異，一度被誤認為密碼根本不存在。
Cyphral Distich是什麼？
是蘇格蘭作家Thomas Urquhart在1653年著作《Logopandecteision》書末的64位數字密碼，自1899年起被學界視為未解之謎。
Fable 5.1花多久、用了多少token解出密碼？
44分鐘、176,000個token，整個搜尋與驗證過程沒有人類介入。
這件事對一般AI使用者有什麼意義？
提醒大家對AI公司或評測機構發布的「破紀錄」跑分新聞稿要保持懷疑，結論要回頭核對原始資料才算數，不要只看標題就轉發。
延伸閱讀 / Related Articles
- 華為Ascend 960評測：提前9個月出貨，靠堆量硬拚輝達 | Huawei Ascend 960 Review: Ships Early, Skips EUV to Rival Nvidia
- Qwen3.8-27B評測：RTX 5090本地跑多快？實測148 tok/s | Qwen3.8-27B Review: RTX 5090 Hits 148 tok/s Locally
- Siri Gemini評測：iOS 27真登場，10億美元換來什麼 | Siri Gemini Review: iOS 27 Ships, What $1B Buys You
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言