Claude Fable 5.1評測：快取砍75%，資安誤報少六成 | Claude Fable 5.1 Review: Cache Cut 75%, Fewer False Flags

By Kit 小克 | AI Tool Observer | 2026-09-03

🇹🇼 Claude Fable 5.1評測：快取砍75%，資安誤報少六成

Claude Fable 5.1 與 Mythos 5.1 在 2026 年 9 月 1 日由 Anthropic 正式推出，這是 Fable 5 系列的第一次小改版，重點不在跑分屠榜，而是把「用起來的痛點」一次修掉：快取讀取費砍 75%、資安工具誤報砍 60%、生物安全過濾誤報砍 85%。對每天靠 Claude Code 幹活的開發者來說，這次更新比起分數更新更有感。

Fable 5.1 跟 Mythos 5.1 差在哪？

先說清楚一個常被搞混的地方：Fable 5.1 跟 Mythos 5.1 是同一個模型、同一組權重，差別只在安全防護層的設定。Fable 5.1 對所有人開放，Mythos 5.1 則是限「受信任存取計畫」的美國組織才能用，主要拿掉了對資安研究、生物科技研究等敏感領域的保守限制，讓真的有正當需求的研究單位不用一直被模型「已讀不回」。

跑分：Terminal-Bench-Science 直接翻倍

硬跑分數字確實漂亮：Terminal-Bench-Science 從 24.7% 衝到 52.6%，幾乎是雙倍成長；一般編碼用的 Terminal-Bench 4.0 從 42.0% 提升到 55.8%；SWE-bench Verified 拿下 95.0%，形式數學證明的 ProofBench v1.1 更是滿分 100%。不過要提醒一句：這些是廠商自己公布的跑分，實際專案的體感表現還是得自己跑過才準。

對開發者最實際的三件事

快取讀取從每百萬 token 1 美元砍到 0.25 美元，官方估算一般工作流成本降 25%，跑 Agent 密集任務甚至可省到 45%
思考模式改成預設開啟，不用再手動開 extended thinking，用 effort 參數（low/medium/high/xhigh/max）五段式控制思考深度
Debug 行為改變：官方跟第三方測試都提到，Fable 5.1 比較不會貼 OK 繃式的快速修法，會傾向真的把 root cause 挖出來，這對維運型任務是好消息但也代表單次回應可能更花時間

企業端：EFS 讓資料零保留也能兼顧安全監控

Anthropic 同步推出 Enterprise Frontier Safeguards（EFS），讓企業可以在自己控制的雲端基礎架構上做到「零資料保留」，同時安全監控機制照樣運作。這對金融、醫療這類對資料主權特別敏感的產業，是一個實際能用的折衷方案。

價格沒變，但要留意什麼

輸入/輸出 token 價格維持 每百萬 10 美元 / 50 美元，跟 Fable 5 一樣，是 Opus 5（5 美元/25 美元）的兩倍。定價沒漲是好消息，但也代表 Anthropic 把 Fable 5.1 定位在「旗艦但不是最貴」的位置——真正省錢的地方在快取，如果你的工作流沒有大量重複讀取上下文，實際省下的錢會比官方宣稱的數字少。Hacker News 上的討論熱度不算特別高（單篇貼文僅 67 分、留言寥寥），社群普遍反應是「穩健升級」而非「驚艷突破」，實測資料目前也還不多，建議自己拿現有專案跑一輪再決定要不要換版本。

好不好用，試了才知道。

🇺🇸 Claude Fable 5.1 Review: Cache Cut 75%, Fewer False Flags

Claude Fable 5.1 and Mythos 5.1 launched on September 1, 2026 — Anthropic's first point release for the Fable 5 line. This isn't a benchmark-chasing update; it's a fix for the papercuts developers actually complain about: a 75% cut to cache-read pricing, 60% fewer cybersecurity false positives, and 85% fewer biosafety filter false positives. For anyone running Claude Code daily, this update matters more in practice than on a leaderboard.

What's the difference between Fable 5.1 and Mythos 5.1?

First, clear up the confusion: Fable 5.1 and Mythos 5.1 are the exact same model and weights — the only difference is the safeguard configuration. Fable 5.1 is generally available to everyone; Mythos 5.1 is restricted to trusted-access program participants (US organizations only, for now), with looser guardrails around sensitive domains like cybersecurity and life-sciences research, so legitimate researchers stop getting stonewalled by overcautious refusals.

Benchmarks: Terminal-Bench-Science more than doubled

The raw numbers are strong: Terminal-Bench-Science jumped from 24.7% to 52.6%, nearly doubling; the general coding benchmark Terminal-Bench 4.0 rose from 42.0% to 55.8%; SWE-bench Verified hit 95.0%; and ProofBench v1.1, for formal math proofs, scored a perfect 100%. One caveat: these are vendor-reported numbers — always worth validating against your own workload before trusting them.

Three things that actually matter for developers

Cache reads dropped from $1.00 to $0.25 per million tokens — Anthropic estimates roughly 25% lower cost for typical workloads, and up to 45% for agentic, context-heavy tasks
Extended thinking is now on by default, no longer opt-in, controlled via a five-level effort parameter (low/medium/high/xhigh/max)
Debugging behavior changed — both Anthropic and early testers report Fable 5.1 is less likely to slap on a quick band-aid fix, and more likely to actually hunt down the root cause. Good news for maintenance work, but expect individual responses to take longer

Enterprise: EFS lets you keep zero data retention and safety monitoring

Anthropic also shipped Enterprise Frontier Safeguards (EFS), letting organizations run zero data retention on their own controlled cloud infrastructure while safety monitoring keeps functioning. That's a practical middle ground for finance and healthcare teams who care deeply about data sovereignty but still need safety guarantees.

Pricing held steady — but read the fine print

Input/output pricing stays at $10 / $50 per million tokens, unchanged from Fable 5 and double Opus 5's $5/$25. No price hike is good news, but it also confirms Anthropic is positioning Fable 5.1 as "flagship, not cheapest" — the real savings are in caching, and if your workflow doesn't reuse a lot of context, you'll see less benefit than the headline numbers suggest. The Hacker News launch thread stayed fairly quiet (just 67 points, one comment), and early community sentiment reads as "solid, incremental upgrade" rather than "breakthrough." Real-world usage data is still thin — worth running it against your own project before switching over.

You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

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

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士