Fable 5 最新動態與時間線

完整時間線：從發布到停用

根據公開來源獨立整理。官方聲明、媒體報導、社群討論和預測會分開標註；每條都連到來源。

1. 2026 年 1 月 報導

   Anthropic 阻止第三方工具使用訂閱登入

   伺服器端檢查讓 OpenCode 等工具無法再用 Claude Pro/Max 訂閱 token 登入；Anthropic 稱這違反服務條款。

   The Register →
2. 2026 年 2 月 報導

   消費者訂閱登入被限制在 Claude 官方應用內

   Anthropic 進一步說明，Free/Pro/Max 登入 token 不能用於其他產品，包括 Agent SDK；OpenCode 隨後移除 Claude 支援。

   Winbuzzer →
3. 2026 年 4 月 報導

   訂閱用戶被告知第三方工具需額外付費

   Anthropic 向用戶發信稱，訂閱額度不能覆蓋 OpenClaw 等第三方工具調用，並一度切斷 OpenClaw 作者的付費 API 存取。

   TechCrunch →
4. 2026 年 5 月 13 日 報導

   Agent SDK / 程式化使用的單獨額度池被宣布

   Anthropic 稱從 6 月 15 日起，付費方案會為 Agent SDK、「claude -p」和第三方應用提供獨立月度額度；互動式 Claude Code 不受影響。

   The New Stack →
5. 2026 年 6 月 9 日 官方

   Anthropic 公開發布 Claude Fable 5 和 Mythos 5

   Fable 5 作為 Anthropic 當時最強的公開模型發布，同時推出面向受審查組織的 Mythos 5。

   CNBC →
6. 2026 年 6 月 10 日 社群

   越獄者 “Pliny the Liberator” 聲稱繞過 Fable 5

   一則病毒式傳播的「越獄警報」聲稱繞過了 Fable 5 的安全保護；之後政府提到的風險背景與此類漏洞相關。

   X (@elder_plinius) →
7. 2026 年 6 月 12 日 官方

   美國出口管制指令迫使 Fable 5 和 Mythos 5 下線

   Anthropic 稱收到指令，需暫停任何外國公民存取；由於無法即時核驗國籍，公司對全球所有客戶關閉兩個模型。Opus 4.8 等其他模型不受影響。

   Anthropic →
8. 2026 年 6 月 12 日 報導

   商務部長信件據稱提到向中俄轉移風險

   公開報導顯示，美國商務部長 Howard Lutnick 致信 Dario Amodei，稱存在向受關注國家軍事用戶轉移的「不可接受風險」。

   Bloomberg →
9. 2026 年 6 月 12 日 報導

   媒體稱 Amazon 的警告觸發了行動

   報導稱，Amazon CEO Andy Jassy 向政府提示 Fable 5 安全問題後，相關行動開始；Anthropic 據稱只有約 90 分鐘下線模型。

   Fortune →
10. 2026 年 6 月 13 日 官方

    Anthropic 反駁召回必要性和越獄嚴重性

    Anthropic 稱不同意召回必要性，認為演示只是「狹窄、非普遍」的越獄，暴露的是較小且已知的問題，並稱正在努力恢復存取。

    Fortune →
11. 2026 年 6 月 13 日 報導

    美國首次把出口管制直接用於 AI 模型

    報導將此事描述為美國出口管制首次直接作用於 AI 模型，而不是晶片；David Sacks 等政府相關人士也參與公開討論。

    Al Jazeera →
12. 2026 年 6 月 14 日 報導

    Sacks 稱政府要求 Anthropic 修復或下架

    白宮 AI 顧問 David Sacks 表示，一旦越獄問題被修復，限制有可能解除，並稱「球在 Anthropic 一邊」。

    Tom's Hardware →
13. 2026 年 6 月 14 日 報導

    Anthropic 派高階員工赴華盛頓密集溝通

    Axios 報導稱，Anthropic 派技術人員前往華盛頓，並幾乎每天與官員會面，試圖推動解除禁令。

    Axios →
14. 2026 年 6 月 15 日 官方

    原計畫的 Agent SDK 額度調整被暫停

    Anthropic 暫停單獨的程式化使用額度。幫助中心稱目前沒有變化：Agent SDK、「claude -p」和第三方用量仍計入訂閱原有限額。

    Claude Help Center →
15. 2026 年 6 月 16 日 預測

    預測市場押注相對較快恢復

    Kalshi 交易者給出約 58% 的概率，認為存取會在 7 月 1 日前恢復。這反映的是市場情緒，不是分析師預測。

    CNBC →
16. 2026 年 6 月 17 日 官方

    Fable 5 和 Mythos 5 仍暫停，暫無恢復日期

    截至 6 月中旬，兩個模型仍未恢復，也沒有承諾的恢復日期；Anthropic 稱仍在努力恢復存取，談判繼續。

    Anthropic →
17. June 17, 2026 官方

    Anthropic exec: models back “in the coming days”

    At a Seoul press conference, Anthropic's Managing Director of International Chris Ciauri says the company is “very confident that in the coming days, the models will become available again.” By late June that had not happened.

    Korea JoongAng Daily →
18. June 18, 2026 報導

    Reported trigger: a South Korean telecom's early Mythos access

    WIRED reports the directive was prompted in part by SK Telecom — a $100M Anthropic investor flagged over suspected China ties — gaining early Mythos 5 access through Anthropic's Project Glasswing consortium.

    WIRED →
19. June 19, 2026 報導

    Trump softens: Anthropic not a security threat “now”

    Asked on The Axios Show whether he views Anthropic as a national-security threat, President Trump says “not now, but a week ago, maybe,” and calls Dario Amodei “nice” and “smart” after meeting him at the G7 — while not ruling out Defense Production Act powers.

    Axios →
20. June 24, 2026 報導

    Reports: Amodei pulled from White House talks, Tom Brown steps in

    Per WIRED-sourced reporting, Anthropic sidelined CEO Dario Amodei from the negotiations — handing them to Anthropic's Tom Brown and policy chief Sarah Heck — with one official quoted that Brown “can actually engage.” This is about the talks only; there is no evidence Amodei is leaving as CEO.

    Gizmodo →
21. June 25, 2026 官方

    A viral “Fable 5 is back” wave is a UI bug — access still zero

    After screenshots claim Fable 5 reappeared in Claude Code's model picker, Anthropic staff say they are “serving exactly 0 traffic” to the model and call reports of restored access “categorically false” — a front-end display bug, not real access.

    explainx.ai →
22. June 26, 2026 發布

    OpenAI launches GPT-5.6 (Sol / Terra / Luna) — also gated by the US government

    OpenAI previews a tiered GPT-5.6 family; flagship Sol claims a Terminal-Bench lead over Claude Mythos 5 and the suspended Fable 5 (OpenAI-reported numbers). Like Fable 5, it is access-restricted: only ~20 government-approved partners can use it, via the API and Codex — not consumer ChatGPT.

    Axios →
23. June 26, 2026 官方

    The first model is cleared to return: Mythos 5 for 100+ US organizations

    Anthropic says the US government notified it that Mythos 5 — its strongest cybersecurity model and Fable 5's sibling — can be redeployed to US organizations that operate and defend critical infrastructure (reported as 100-plus institutions, and notably including their non-American employees). Anthropic is restoring access for them quickly and says it's still working to make Fable 5 available for general use again. Fable 5 itself remains suspended for the public.

    NBC News →
24. June 27, 2026 發布

    Mythos 5 starts coming back online — on Amazon Bedrock and Microsoft Foundry

    The cleared Mythos 5 begins redeploying to vetted US critical-infrastructure organizations through Anthropic's Project Glasswing, with model access returning on Amazon Bedrock and Microsoft Foundry. Fable 5 for the general public is still not back — fifteen days into the suspension.

    AWS →
25. June 30, 2026 官方

    The US lifts the export controls on Fable 5

    The US government withdraws the June 12 export controls on Claude Fable 5 and Mythos 5, clearing the way for a full public return. Anthropic announces redeployment plans the same day.

    Al Jazeera →
26. July 1, 2026 官方

    Fable 5 returns worldwide — plus a new cheaper model, Claude Sonnet 5

    Anthropic restores Fable 5 globally across the Claude platform, Claude.ai, Claude Code, and Cowork, included on paid plans for up to 50% of weekly usage limits through July 7. Claude Sonnet 5, a cheaper agent-focused model, ships alongside it at $2/$10 per million tokens (through Aug 31).

    Anthropic →
27. July 8, 2026 官方

    OpenAI: GPT-5.6 to go public July 9 — the government gate lifts

    OpenAI confirms that GPT-5.6 Sol, Terra, and Luna launch publicly on Thursday, July 9, after the US Commerce Department cleared a broad release. Announced API pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million tokens — Sol at half of Fable 5's token rate, three days before Fable 5's plan window was then due to close (July 12).

    Neowin →
28. July 9, 2026 官方

    GPT-5.6 goes live for everyone — ChatGPT, API, and Codex

    As announced, OpenAI makes Sol, Terra, and Luna generally available on July 9 with global rollout completing within a day. Pricing confirmed: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million tokens. GitHub Copilot adds all three the same day. The two-week government gate is gone — three days before Fable 5's plan window was then due to close (July 12).

    OpenAI →
29. July 8, 2026 官方

    Extension: Fable 5 stays in paid plans through July 12

    Hours after the window was due to close, Anthropic announces a five-day extension: Fable 5 remains included on all paid plans through July 12, still up to 50% of weekly usage limits, applying automatically. After July 12 it moves to usage credits at $10 input / $50 output per million tokens. Anthropic maintains the credits move isn't permanent and that subscription access returns as capacity allows.

    TestingCatalog on X →
30. July 8, 2026 官方

    xAI ships Grok 4.5 — a cheaper coding model lands in the switching week

    xAI releases Grok 4.5, its first model built for coding and agentic work, priced at $2 input / $6 output per million tokens — well under Fable 5's $10/$50 usage-credit rate. Elon Musk calls it "Opus-class"; independent testing is more measured. Artificial Analysis places it fourth on its Intelligence Index (54), behind Fable 5 (60), Opus 4.8 (56), and GPT-5.5 (55); its Terminal-Bench 2.1 score of 83.3 (xAI's own number) sits just under Fable 5's 84.3. A genuine low-cost option to weigh as Fable 5's plan window nears its close (since extended to July 19).

    TechCrunch →
31. July 12, 2026 官方

    Extended again: Fable 5 stays in paid plans through July 19

    In a late-night post from its official @claudeai account — with the promotional-access support article updated to match — Anthropic extends the plan-included window a second time, now through July 19. Terms unchanged: Pro, Max, Team, and select Enterprise plans, up to 50% of weekly usage limits, applying automatically. After July 19, Fable 5 bills through usage credits at $10 input / $50 output per million tokens. Anthropic continues to say the credits move isn't permanent.

    @claudeai on X →
32. July 18, 2026 官方

    Permanent: Fable 5 stays on Max and Team Premium — Pro moves to credits with $100 included

    One day before the window was due to close, Anthropic announces the lasting arrangement: from July 20, Fable 5 remains included on Max and Team Premium plans at up to 50% of weekly limits — described as permanent, not another extension — while Pro and Team Standard move to usage credits at $10 input / $50 output per million tokens, cushioned by a one-time $100 credit. The reversal lands amid open price pressure from GPT-5.6 Sol at half Fable 5's token rate.

    The Decoder →
33. July 16, 2026 官方

    Moonshot ships Kimi K3: near-Sol coding scores, open weights, $3/$15

    Moonshot AI releases Kimi K3, a 2.8-trillion-parameter MoE with a 1M-token context and native vision. Self-reported Terminal-Bench 2.1: 88.3 — just under GPT-5.6 Sol's 88.8 and above Fable 5's published score — at $3/$15 per million tokens with a 90% cache discount. Open weights are promised by July 27. Another serious budget seat in the model mix as Fable 5's split takes effect.

    Simon Willison →
34. July 24, 2026 官方

    Anthropic ships Claude Opus 5 — near-Fable intelligence at half the price

    Four days after the split took effect, Anthropic launches Claude Opus 5 at $5 input / $25 output per million tokens — unchanged from Opus 4.8, and exactly half of Fable 5's rate. The company positions it as delivering nearly all of Fable 5's intelligence; it becomes the default model on Max and the strongest model included with Pro. Anthropic also says Opus 5 runs with roughly 85% fewer safety-classifier interventions than Fable 5 and is exempt from the 30-day retention policy. For Pro users weighing the one-time $100 Fable credit, this resets the question.

    CNBC →