Claude Fable 5.1評測：快取降價75%，3項破壞性API更新 | Claude Fable 5.1 Review: 75% Cheaper Cache, Breaking Changes

9月 06, 2026

By Kit 小克 | AI Tool Observer | 2026-09-06

🇹🇼 Claude Fable 5.1評測：快取降價75%，3項破壞性API更新

Claude Fable 5.1 是 Anthropic 在 2026 年 9 月 1 日發布的新版旗艦模型，同步上線的還有僅開放信任名單使用的 Claude Mythos 5.1。這次更新最實際的變化不是「更聰明」，而是快取讀取費用砍 75%，以及三項會讓舊程式碼直接壞掉的破壞性 API 變更。如果你的 Agent 或 pipeline 正在用 Claude API，這篇文章值得你花五分鐘看完再上線。

Claude Fable 5.1 到底改了什麼

Fable 5.1 與 Mythos 5.1 其實是同一個底層模型，差別只在安全防護等級：Mythos 5.1 拿掉了針對生物與資安領域的限制，只開放給 Anthropic 的信任存取計畫使用；一般開發者能拿到的是 Fable 5.1。根據 Anthropic 官方公告，Fable 5.1 在 Terminal-Bench-Science 這項評測拿下 52.6%，遠高於前代 Fable 5 的 24.7%，甚至超過 Opus 5（29.0%）與 GPT-5.6 Sol（22.4%）。

價格：基本費率沒變，但快取便宜很多

輸入 / 輸出價格維持不變：每百萬 token 10 美元 / 50 美元
快取讀取從每百萬 token 1 美元降到 0.25 美元，降幅 75%
Anthropic 估算一般工作流成本可省約 25%，Agent 密集型任務最高可省 45%

3 個破壞性變更，upgrade 前務必檢查

編輯歷史會讓 thinking block 失效：只要修改 thinking block 之前的內容（system prompt、tools 陣列、前面的訊息），下一次請求就會報錯或被靜默丟棄，視設定而定。這項檢查適用於 2026 年 8 月 31 日之後建立的新帳號。
Parallel tool calling 行為變得不穩定：原本 Fable 5 一次會批次呼叫多個工具，Fable 5.1 可能改成每輪只呼叫一次，Agent loop 的邏輯要重新測試。
安全防護規則調整，減少過去在生物、資安主題上的誤判擋掉正常請求的狀況，但代表你的 prompt 測試案例可能要重跑一輪。

老實說值不值得升級？

如果你只是單純聊天用途，這次升級感受不大。但如果你在跑大量 prompt caching的 Agent 系統（例如長文件分析、RAG、coding agent），75% 的快取降價是實打實的成本節省，值得盡快遷移。反過來說，如果你的程式碼有依賴「thinking block 之前可以任意編輯歷史」這個舊行為，或是依賴 Fable 5 的批次工具呼叫模式，先在測試環境跑過一輪再上生產，不要直接切版。

好不好用，試了才知道。

🇺🇸 Claude Fable 5.1 Review: 75% Cheaper Cache, Breaking Changes

Claude Fable 5.1 is Anthropic's new flagship model release from September 1, 2026, launched alongside Claude Mythos 5.1, which is restricted to trusted-access participants. The headline news isn't raw intelligence gains — it's a 75% cut to cache read pricing and three breaking API changes that can silently break existing agent code. If you're running production workloads on the Claude API, this is worth five minutes before you upgrade.

What Actually Changed in Claude Fable 5.1

Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguard levels — Mythos 5.1 removes restrictions around cybersecurity and biology topics and is only available through Anthropic's trusted access program, while regular developers get Fable 5.1. Per Anthropic's official announcement, Fable 5.1 scores 52.6% on Terminal-Bench-Science, up sharply from Fable 5's 24.7%, and ahead of both Opus 5 (29.0%) and GPT-5.6 Sol (22.4%).

Pricing: Base Rates Unchanged, Cache Reads Much Cheaper

Input/output pricing stays the same: $10 / $50 per million tokens
Cache reads drop from $1.00 to $0.25 per million tokens — a 75% reduction
Anthropic estimates roughly 25% lower cost on typical workloads, up to 45% on agentic ones

3 Breaking Changes to Check Before You Upgrade

Editing history invalidates thinking blocks: change anything before a thinking block — system prompt, tools array, an earlier message — and the next request either errors or silently drops the block, depending on configuration. This applies to accounts created on or after August 31, 2026.
Parallel tool calling is less predictable: Fable 5 used to batch multiple tool calls per turn; Fable 5.1 may issue only one call per turn instead, which can break agent loop logic that assumed batching.
Safeguard behavior was retuned to reduce false-positive blocks on biology and cybersecurity topics — meaning your existing prompt test suites may need a re-run to confirm nothing regressed elsewhere.

Is It Worth Upgrading? Honestly, Depends

If you're just chatting casually, you won't notice much. But if you're running heavy prompt-caching agent systems — long-document analysis, RAG pipelines, coding agents — the 75% cache discount is a real cost win worth migrating for soon. On the flip side, if your code relies on the old behavior of freely editing history before a thinking block, or depends on Fable 5's batched tool-calling pattern, test in staging first before flipping production traffic over.

You won't know if it's good 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

搜尋此網誌

AI小貼士