Claude Opus 5評測：半價逼近Fable 5，該換嗎？ | Claude Opus 5 Review: Half the Price of Fable 5

7月 30, 2026

By Kit 小克 | AI Tool Observer | 2026-07-31

🇹🇼 Claude Opus 5評測：半價逼近Fable 5，該換嗎？

Claude Opus 5是什麼：Anthropic的「半價旗艦」

Claude Opus 5 是 Anthropic 在 2026 年 7 月 24 日推出的新一代模型，定位介於今年 6 月上市的 Fable 5 與即將退役的 Opus 4.8 之間。最大賣點很直白：價格維持 Opus 4.8 原價（每百萬輸入 token 5 美元、輸出 25 美元），但實測表現卻逼近貴了一倍的 Fable 5（每百萬輸入 10 美元、輸出 50 美元）。目前 Claude Max、Pro 訂閱與 API 都已全面開放使用。

跑分實測：Agentic Coding 是強項

Anthropic 官方數據顯示，Opus 5 在 SWE-bench Pro 拿下 79.2% 的成績，在專門測試長流程軟體工程任務的 Frontier-Bench v0.1 更是拿下 43.3%，直接把 Fable 5 的 33.7% 甩在後面，而且成本更低。在 agentic 任務上，OSWorld 2.0 達 70.57%、AutomationBench 26.0%、ARC-AGI-3 也有 30.16%，這幾項都是需要模型自己規劃多步驟、操作電腦畫面的任務，顯示 Opus 5 的強項確實在「代理型」工作流程，而不是單純的知識問答。

新功能：Effort開關與Fast模式

這次更新加入一個叫「Effort」的參數，讓開發者自己決定模型要花多少運算力去思考，直接對應成本與品質的取捨。另外還有 Fast 模式，速度快 2.5 倍但價格也是雙倍；API 端也內建了自動 fallback 路由，換工具時不會打斷 prompt cache，對正在做長對話 agent 產品的團隊算是實用的細節優化。值得一提的是，Opus 5 的知識截止日是 2026 年 5 月，比 Fable 5 和 Opus 4.8 的 2026 年 1 月更新，這對需要問近期時事或技術的場景有差。

該不該換？誠實建議

如果你原本就在用 Opus 4.8 做程式碼相關工作，這次幾乎是「同價升級」，沒有理由不換。如果你在用 Fable 5 做複雜的多步驟 coding 或電腦操作任務，可以先拿 Opus 5 跑跑看，很可能省下一半費用又拿到差不多的結果。但如果你要的是最頂尖的通用推理與寫作品質，Fable 5 目前還是天花板，Opus 5 沒有要取代它的意思，只是把「夠好」的門檻拉低到更便宜的價位。

好不好用，試了才知道。

🇺🇸 Claude Opus 5 Review: Half the Price of Fable 5

Claude Opus 5: Anthropic's "Half-Price Flagship"

Anthropic launched Claude Opus 5 on July 24, 2026, positioning it as the practical middle ground between June's Fable 5 and the outgoing Opus 4.8. The pitch is simple: it keeps Opus 4.8's original pricing ($5 per million input tokens, $25 per million output tokens) while closing most of the performance gap with Fable 5, which costs twice as much ($10/$50 per million tokens). It's already live across Claude Max, Pro, and the API.

Benchmarks: Strongest at Agentic Coding

According to Anthropic, Opus 5 scores 79.2% on SWE-bench Pro and 43.3% on Frontier-Bench v0.1 — a benchmark for long-horizon software engineering tasks — well ahead of Fable 5's 33.7%, at a lower cost per task. On agentic evaluations it posts 70.57% on OSWorld 2.0, 26.0% on AutomationBench, and 30.16% on ARC-AGI-3. These all measure multi-step planning and computer-use tasks, which is where Opus 5's real strength shows — not raw trivia knowledge, but getting things done across a long, messy workflow.

New Features: Effort Dial and Fast Mode

This release introduces an "Effort" setting that lets developers explicitly trade off compute against quality. There's also a Fast mode running at 2.5x speed for double the price, plus automatic fallback routing built into the API and support for swapping tools mid-conversation without breaking the prompt cache — a small but genuinely useful detail if you're running long-lived agent sessions. Worth noting: Opus 5's training data cutoff is May 2026, fresher than the January 2026 cutoff on both Fable 5 and Opus 4.8.

Should You Switch? Honest Take

If you're already on Opus 4.8 for coding work, this is essentially a free upgrade — same price, better results, no reason not to switch. If you're paying Fable 5 prices for complex multi-step coding or computer-use tasks, it's worth benchmarking Opus 5 against your actual workload — you may get comparable output at half the cost. But if you need the absolute ceiling on general reasoning and writing quality, Fable 5 is still the top model; Opus 5 isn't trying to replace it, it's lowering the price of "good enough."

好不好用，試了才知道。

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