Fable 5.1評測：快取讀取降75%，代理工作省45% | Fable 5.1 Review: Cache Reads Cut 75%, Agents Save 45%

9月 14, 2026

By Kit 小克 | AI Tool Observer | 2026-09-15

🇹🇼 Fable 5.1評測：快取讀取降75%，代理工作省45%

Anthropic 在 9 月 1 日把 Fable 5.1 正式推上生產環境，這是目前 API 市場討論度最高的話題：不是因為它跑分破紀錄，而是因為快取讀取費用直接砍 75%，從每百萬 token 1 美元降到 0.25 美元。對每天跑長對話、多輪 agent 任務的開發者來說，這比模型變聰明更有感。

Fable 5.1 改了什麼

Fable 5.1 是 Fable 5 的小改款，輸入輸出價格維持每百萬 token 10 美元 / 50 美元沒變，但這次升級鎖定六個實際場景：

長時間 agentic coding：跨檔案重構、拖好幾小時的除錯任務
文件與試算表產出：從空白檔案直接做出完整簡報或報表
多步驟研究：能追問自己剛查到的結果，不用每次重新下指令
密集文件的視覺理解：讀 PDF 裡塞滿的圖表與附件
完整運用 1M token 視窗推理，不是塞得進去就好，是塞進去還能用
電腦操作恢復力：步驟失敗後更會自己接回來

根據 MarkTechPost 的測試，Fable 5.1 在 Terminal-Bench-Science 拿到 52.6%，代表它在需要多步驟操作終端機、跑實驗的任務上比前代更穩。

快取降價 75%，實際省多少？

Anthropic 官方公布的定價文件寫得很直白：一般用量大概省 25%，但如果是重度 agentic 工作流（反覆讀同一份系統提示詞、長上下文對話），最多可以省到 45%。這個數字對做 agent 產品的團隊特別關鍵，因為快取讀取通常佔了帳單很大一塊，不是主要輸入輸出 token。

老實說：值不值得換？

The New Stack 的實測提到一個容易被忽略的細節：Fable 5.1「拒絕回答」的次數明顯變少，過去常見的誤判安全問題有改善，這對做客服或內容產品的團隊是實際痛點。另外 Anthropic 同步推出 Enterprise Frontier Safeguards（EFS），資料存在客戶自己控制的雲端基礎設施，不經過 Anthropic，對重視資料主權的企業客戶是加分。

如果你的應用是短對話、低快取命中率，這次升級對帳單影響有限；但只要你在跑長上下文 agent 任務，這次降價幅度值得馬上測試看看。

好不好用，試了才知道。

🇺🇸 Fable 5.1 Review: Cache Reads Cut 75%, Agents Save 45%

Fable 5.1 is the AI pricing story developers are actually talking about this week — not because it broke a benchmark record, but because Anthropic cut cache read pricing by 75%, from $1.00 to $0.25 per million tokens. For anyone running long-context agents or multi-turn coding sessions, that matters more than a few extra benchmark points.

What Actually Changed in Fable 5.1

Fable 5.1 is an incremental update to Fable 5. Input/output pricing stays the same at $10/$50 per million tokens, but Anthropic targeted six concrete workflows this time:

Long-session agentic coding — multi-file refactors and hours-long debugging runs
Document and spreadsheet work — going from a blank page to a finished deliverable
Multistep research that follows up on its own findings instead of waiting for a new prompt
Vision on dense documents — charts and filings buried inside PDFs
Reasoning across the full 1M-token context window, not just fitting text into it
More reliable computer use, with better recovery after a failed step

According to MarkTechPost's benchmark coverage, Fable 5.1 scored 52.6% on Terminal-Bench-Science — a meaningful jump for tasks that require multi-step terminal operations.

How Much Does the 75% Cache Cut Actually Save?

Per Anthropic's own pricing docs, typical workloads get roughly 25% cheaper overall, while heavily agentic workflows — the ones that keep re-reading the same system prompt or long conversation history — can save up to 45%. That matters because cache reads, not raw input/output tokens, often dominate the bill for agent products.

Honest Take: Is It Worth Switching?

The New Stack's hands-on review flagged something easy to miss: Fable 5.1 refuses far fewer legitimate requests, fixing a real pain point for teams building customer-facing or content products. Anthropic also shipped Enterprise Frontier Safeguards (EFS) alongside it, keeping data in infrastructure the customer controls rather than Anthropic's own cloud — a plus for enterprises with strict data residency needs.

If your app runs short, low-cache-hit conversations, this upgrade won't move your bill much. But if you're running long-context agents, the pricing cut alone is worth testing this week.

好不好用，試了才知道 — the only way to know if it is worth it is to try it yourself.

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

搜尋此網誌

AI小貼士