Anthropic Releases New AI Models
Anthropic Releases New AI Models
Anthropic 發佈新 AI 模型
Introduction
Anthropic has two new AI models. Claude Fable 5 is for everyone. Claude Mythos 5 is only for a few special partners.
Anthropic 推出了兩款新的 AI 模型。Claude Fable 5 面向所有使用者，而 Claude Mythos 5 僅限於少數特定合作夥伴。
Main Body
Claude Fable 5 is safe. If a user asks a dangerous question about science or computers, the AI stops. It then uses an older, simpler model instead.
Claude Fable 5 非常安全。如果使用者詢問關於科學或電腦的危險問題，AI 會停止運作，隨後改用較舊且較簡單的模型。
Claude Mythos 5 is for 200 groups and the US government. These experts use it for science and computer security. Anthropic saves all their data for 30 days to keep the system safe.
Claude Mythos 5 供 200 個團體及美國政府使用。這些專家將其用於科學研究與電腦安全。Anthropic 會將所有數據保存 30 天，以確保系統安全。
These new models cost more money than the old ones. However, big companies like Rakuten say the AI works much better. Anthropic is now worth 965 billion dollars.
這些新模型的費用比舊版更高。然而，像樂天（Rakuten）這樣的大公司表示，該 AI 的運作效果提升許多。Anthropic 目前的估值為 9,650 億美元。
Conclusion
Anthropic now gives its best AI to the public and partners. They use high prices and safety rules to manage the new tools.
Anthropic 現在將其最強的 AI 提供給公眾與合作夥伴。他們透過高定價與安全規範來管理這些新工具。
Vocabulary Learning
Comparing Things
In the text, we see two ways to say something is 'more' than something else. This is a key skill for A2 learners.
1. Using 'More' with long words When a word is long (like expensive), we put more in front of it.
- Old models New models cost more money
2. Using '-er' with short words When a word is short (like old), we add -er to the end.
- Old models Older, simpler model
Quick Logic Map:
Short Word add -er (e.g., Fast Faster)
Long Word add more (e.g., Powerful More powerful)
Vocabulary Learning
Anthropic Releases New Claude Fable 5 and Mythos 5 AI Models
Anthropic 發佈全新 Claude Fable 5 與 Mythos 5 AI 模型
Introduction
Anthropic has released Claude Fable 5 for the general public and Claude Mythos 5 for a small group of trusted partners. This marks the first time the company has made its most advanced model architecture widely available.
Anthropic 已向大眾發佈 Claude Fable 5，並向少數受信任的合作夥伴提供 Claude Mythos 5。這標誌著該公司首次將其最先進的模型架構廣泛開放。
Main Body
The release follows a preview version of the Mythos model, which was known for its ability to find serious software bugs. To prevent people from using the AI for harmful purposes in fields like cybersecurity, biology, and chemistry, Anthropic added safety filters to Claude Fable 5. If a user asks a dangerous question, the system automatically switches the request to the older Claude Opus 4.8 model. Anthropic emphasized that after more than 1,000 hours of safety testing, about 95% of sessions stayed on the Fable 5 model without needing to switch.
此次發佈繼 Mythos 模型的預覽版之後，該版本以能發現嚴重的軟體漏洞而聞名。為了防止人們將 AI 用於網路安全、生物學和化學等有害目的，Anthropic 在 Claude Fable 5 中加入了安全過濾機制。如果使用者提出危險問題，系統會自動將請求切換至舊版的 Claude Opus 4.8 模型。Anthropic 強調，經過 1,000 多小時的安全測試，約 95% 的對話仍維持在 Fable 5 模型而無需切換。
At the same time, Claude Mythos 5 is available to a limited group called Project Glasswing, which includes 200 organizations and the U.S. government. This restricted access allows researchers and security experts to use the model's full power to create new scientific theories and find system flaws. Furthermore, the company has introduced a mandatory 30-day data storage rule for all users to protect against new types of attacks, even for clients who previously had no data stored.
與此同時，Claude Mythos 5 僅向名為 Project Glasswing 的限定小組開放，其中包括 200 個組織及美國政府。這種受限的訪問權限允許研究人員和安全專家利用模型的全部能力來創造新的科學理論並發現系統缺陷。此外，公司為所有使用者引入了強制性的 30 天數據儲存規則，以防範新類型的攻擊，即使是先前未儲存數據的客戶亦然。
From a business perspective, both new models cost twice as much as Opus 4.8, with prices set at $10 per million input tokens and $50 per million output tokens. Despite the higher cost, companies like Rakuten and Base44 reported that the models provide a better return on investment because they are much better at autonomous tasks and software engineering. This rollout happens as Anthropic prepares for a public stock offering (IPO) and reaches a reported value of $965 billion.
從商業角度來看，這兩款新模型的成本均為 Opus 4.8 的兩倍，價格設定為每百萬輸入 token 10 美元，每百萬輸出 token 50 美元。儘管成本較高，但 Rakuten 和 Base44 等公司報告稱，由於模型在自主任務和軟體工程方面的表現大幅提升，因此提供了更好的投資回報。此次推出正值 Anthropic 準備首次公開募股 (IPO) 且據報估值達到 9,650 億美元之際。
Conclusion
Anthropic has moved its most powerful AI from a private preview to a tiered system for the public and partners, combining strict safety measures with a premium pricing strategy.
Anthropic 將其最強大的 AI 從私人預覽轉為面向大眾與合作夥伴的分層系統，將嚴格的安全措施與高端定價策略相結合。
Vocabulary Learning
🚀 The 'B2 Logic' Shift: Moving from Simple to Complex Connections
At the A2 level, you usually connect ideas with and, but, or because. To reach B2, you need to use Logical Connectors that show a professional relationship between two ideas.
Look at how this text moves beyond basic English:
🛠️ The Transition Toolkit
| Instead of A2 (Simple) | Use B2 (Advanced) | Example from Text |
|---|---|---|
| And / Also | Furthermore | "Furthermore, the company has introduced a mandatory 30-day data storage rule..." |
| But | Despite | "Despite the higher cost, companies... reported a better return on investment." |
Why this matters:
- Furthermore tells the reader: "I am adding a new, important point to my argument."
- Despite tells the reader: "I am acknowledging a problem (high cost), but the result is still positive."
💡 Pro-Tip: The 'Despite' Structure
A2 students often say: "The cost is high, but it is good." B2 students say: "Despite [Noun/Phrase], [Main Clause]."
Example: Despite + the higher cost (The Problem) companies reported a better return (The Surprise/Result).
🔍 Vocabulary Upgrade: From 'General' to 'Precise'
Stop using "good" or "bad." Notice these B2-level replacements in the article:
- ❌ Good for business ✅ Better return on investment
- ❌ Available for some ✅ Restricted access
- ❌ Danger ✅ System flaws / Harmful purposes
Vocabulary Learning
Anthropic Implements Stratified Deployment of Mythos-Class Large Language Models
Anthropic 實施 Mythos 級大型語言模型分層部署
Introduction
Anthropic has launched Claude Fable 5 for general public access and Claude Mythos 5 for a restricted cohort of trusted partners, marking the first broad availability of its most advanced model architecture.
Anthropic 已推出面向大眾的 Claude Fable 5 以及面向少數受信任合作夥伴的 Claude Mythos 5，標誌著其最先進的模型架構首次大規模開放。
Main Body
The deployment follows the April introduction of the Mythos Preview, a model characterized by its capacity to identify critical software vulnerabilities. To mitigate the risk of malicious exploitation in cybersecurity, biology, and chemistry, Anthropic has engineered Claude Fable 5 with integrated safety classifiers. Should a query trigger these guardrails, the system automatically reroutes the request to the less capable Claude Opus 4.8 model. Anthropic reports that internal and external red-teaming, exceeding 1,000 hours of testing, yielded no universal jailbreaks, with approximately 95% of sessions remaining on the Fable 5 architecture without fallback.
此次部署延續了 4 月推出的 Mythos Preview，該模型的特點在於能識別關鍵的軟體漏洞。為了降低在網路安全、生物學與化學領域被惡意利用的風險，Anthropic 在 Claude Fable 5 中整合了安全分類器。若查詢觸發這些防護機制，系統會自動將請求重新定向至功能較弱的 Claude Opus 4.8 模型。Anthropic 報告指出，經過超過 1,000 小時的內部與外部紅隊測試，未發現通用的越獄方法，約 95% 的對話維持在 Fable 5 架構而無需回退。
Parallel to the public release, Claude Mythos 5 is available to a limited consortium known as Project Glasswing, which includes approximately 200 organizations across 15 countries and the U.S. government. This restricted access allows cybersecurity professionals and life sciences researchers to utilize the model's full capabilities, including the generation of novel scientific hypotheses and the identification of systemic software flaws. The company has further instituted a mandatory 30-day data retention policy for all traffic to defend against novel attacks, notwithstanding prior zero-retention agreements with enterprise clients.
與公開發佈平行，Claude Mythos 5 向名為 Project Glasswing 的有限聯盟開放，其中包括分布於 15 個國家約 200 個組織以及美國政府。這種限制性存取讓網路安全專業人員與生命科學研究員能利用該模型的全部能力，包括生成新型科學假設與識別系統性軟體缺陷。公司進一步對所有流量實施了強制性 30 天數據保留政策以抵禦新型攻擊，即便先前與企業客戶簽署過零保留協議亦然。
From a commercial perspective, both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens, representing a 100% increase over Opus 4.8. Despite this premium, early institutional feedback from entities such as Rakuten and Base44 suggests a higher return on investment due to superior performance in autonomous operations and software engineering. This strategic rollout coincides with Anthropic's confidential IPO filings and a reported valuation of $965 billion, reflecting a competitive effort to demonstrate technical superiority ahead of public market entry.
從商業角度來看，Fable 5 與 Mythos 5 的定價均為每百萬輸入標記（token）10 美元及每百萬輸出標記 50 美元，較 Opus 4.8 增加 100%。儘管價格較高，但來自 Rakuten 和 Base44 等機構的早期回饋顯示，由於在自動化運作與軟體工程方面表現優異，投資報酬率更高。此次策略性發佈正值 Anthropic 秘密提交 IPO 申請且據報估值達 9,650 億美元之際，反映出其在進入公開市場前力求證明技術領先地位的競爭努力。
Conclusion
Anthropic has transitioned its most powerful AI capabilities from a restricted preview to a tiered public and partner-based availability model, balanced by stringent safety fallbacks and a premium pricing structure.
Anthropic 已將其最強大的 AI 能力從限制性預覽轉型為分層的公開與合作夥伴可用模式，並透過嚴格的安全回退機制與溢價定價結構達成平衡。
Vocabulary Learning
The Architecture of "Nuanced Limitation"
To bridge the gap from B2 to C2, a student must move beyond simple opposites (e.g., safe vs dangerous) and master mitigating qualifiers and concessive logical pivots. In this text, the most potent C2 linguistic phenomenon is the use of formal constraints to signal high-level precision.
⚡ The Pivot: Notwithstanding
Look at the phrase: "...notwithstanding prior zero-retention agreements..."
At B2, a student uses "despite" or "although." At C2, notwithstanding functions as a preposition that allows the writer to acknowledge a contradictory fact without letting it derail the primary clause. It creates a legalistic, authoritative tone that signals the priority of the current action over previous promises.
⚖️ Precision via "Stratified" Lexis
Note the movement from general terms to specific, high-density descriptors:
- B2 Level: Different levels of access
- C2 Level: Stratified deployment / Tiered availability
By using stratified, the author isn't just saying there are "levels"; they are invoking a sociological/geological metaphor for layers, implying a structural, intentional hierarchy. This is the essence of C2: choosing a word that carries an implicit conceptual framework.
🛠️ The "Fallback" Mechanism (Syntactic Compression)
Observe: "...yielding no universal jailbreaks, with approximately 95% of sessions remaining on the Fable 5 architecture without fallback."
The use of the absolute construction ("with approximately 95%... remaining") avoids the clunkiness of "and 95% of sessions remained." This compression allows the writer to maintain the flow of a complex sentence while layering secondary data.
C2 Takeaway: To achieve mastery, stop writing sequences of simple sentences. Start using with + noun + participle to append circumstantial information to your main claim.