GPT-6 Sol/Luna評測：砍半價格，編碼分數不進反退 | GPT-6 Sol & Luna Review: Half Price, Same Coding Bugs

9月 27, 2026

By Kit 小克 | AI Tool Observer | 2026-09-27

🇹🇼 GPT-6 Sol/Luna評測：砍半價格，編碼分數不進反退

GPT-6 Sol與Luna是什麼：Astra的平價版

OpenAI 在 GPT-6 Astra 上市僅 19 天後，於 2026 年 9 月 22 日推出 GPT-6 Sol 與 GPT-6 Luna 這兩款輕量模型，主打「不用旗艦價也能跑 agent 流程」。這次更新的重點不是刷新智力天花板，而是把大量工具呼叫、多輪推理、生產環境寫程式這類「跑量」場景的成本壓下來。

API 價格砍半，是永久調降不是促銷

GPT-6 Sol：每百萬 token 輸入 $2、輸出 $10（原本 GPT-5.6 Sol 是 $4／$20）
GPT-6 Luna：每百萬 token 輸入 $0.10、輸出 $0.50（原本是 $0.20／$1.20）
快取讀取再打一折：Sol 快取輸入只要 $0.20／百萬 token，Luna 只要 $0.01

OpenAI 特別強調這是「永久定價」而非限時優惠，等於把 Claude Opus 5.5 等級的價格帶直接複製過來，兩家的中階模型現在幾乎打平。

老實說：編碼分數沒有變好，反而退步

這是這次評測最誠實、也最需要提醒讀者的地方。根據多家外部測試，在三項編碼與電腦操作基準測試中，有兩項 GPT-6 Sol 的最佳分數其實比舊版 GPT-5.6 Sol 還低。換句話說，如果你把它當成「更聰明的寫程式模型」升級，實測結果可能會讓你失望。

OpenAI 自己拿出的亮點是「事實錯誤率減半」——用真實 ChatGPT 對話中被使用者標記錯誤的案例做內部評測，GPT-6 Sol 的犯錯率大約只有 GPT-5.6 Sol 的一半，官方說法是「逼近 Astra 等級的可靠度」。而 GPT-6 Luna 開到高算力模式時，事實準確度可以打平 GPT-5.6 Sol，但成本只要百分之一。

誰該用、誰該等

如果你的產品是大量 agent 呼叫、需要跑幾千次工具調用的自動化流程，GPT-6 Sol／Luna 的降價加上更省的快取費率，帳單差異會很明顯，值得換。但如果你的核心需求是「寫程式能力要更強」，這次的 Sol 版本不保證比舊版好，實際上兩項基準還退步，建議先用自己的測試集跑過一輪，別只看官方通稿就升級。API model ID 分別是 gpt-6-sol 與 gpt-6-luna，已開放給 ChatGPT、Codex 與 API。

好不好用，試了才知道。

🇺🇸 GPT-6 Sol & Luna Review: Half Price, Same Coding Bugs

What Are GPT-6 Sol and Luna: The Budget Astra Lineup

Just 19 days after launching flagship GPT-6 Astra, OpenAI shipped two lighter models on September 22, 2026: GPT-6 Sol and GPT-6 Luna. The pitch isn't a smarter model — it's making sustained agentic workflows (thousands of tool calls, multi-turn reasoning, production coding) cheaper to run.

API Prices Cut in Half — Permanently

GPT-6 Sol: $2 / $10 per million input/output tokens (down from $4 / $20 for GPT-5.6 Sol)
GPT-6 Luna: $0.10 / $0.50 per million tokens (down from $0.20 / $1.20)
Cached reads get a 90% discount: $0.20/M on Sol, $0.01/M on Luna

OpenAI says these are permanent prices, not a launch promo — putting GPT-6 Sol roughly in the same price tier as Claude Opus 5.5.

The Honest Part: Coding Scores Actually Dropped

This is the part worth flagging before you upgrade anything. Across three coding and computer-use benchmarks tracked by outside reviewers, GPT-6 Sol's best score is lower than GPT-5.6 Sol's on two of them. If you're expecting a straight coding upgrade, real-world tests may disappoint you.

OpenAI's actual highlight is factuality: using an internal eval built from de-identified ChatGPT conversations where users flagged errors, the company says GPT-6 Sol makes roughly half as many mistakes as its predecessor — "approaching Astra-level reliability." At higher effort settings, GPT-6 Luna reportedly matches GPT-5.6 Sol's factuality at about 1% of the task cost.

Who Should Upgrade — and Who Should Wait

If your workload is agent-heavy — thousands of tool calls, automation pipelines, high token volume — the price cut plus cheaper caching on GPT-6 Sol and Luna will show up clearly on your bill, and it's worth switching. But if you specifically need better coding capability, this release doesn't deliver that on paper — two benchmarks actually regressed. Run your own eval set before migrating instead of trusting the announcement alone. Model IDs are gpt-6-sol and gpt-6-luna, available now in ChatGPT, Codex, and the API.

好不好用，試了才知道 — 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

搜尋此網誌

AI小貼士