GPT-6 Sol Luna評測：降價腰斬，效能打幾折 | GPT-6 Sol & Luna Review: Half Price, Same Power?

By Kit 小克 | AI Tool Observer | 2026-09-24

🇹🇼 GPT-6 Sol Luna評測：降價腰斬，效能打幾折

GPT-6 Sol Luna是什麼？

GPT-6 Sol與GPT-6 Luna是OpenAI在2026年9月22日發布的兩款API模型，用同一套訓練方法把旗艦Astra的推理能力下放到便宜檔次，並把價格砍半訂為永久價，不是促銷。這是繼Astra之後，OpenAI第一次認真在「便宜」這件事上動真格。

GPT-6 Sol Luna定價怎麼算？

三層定價現在長這樣：

Astra：每百萬token輸入 $10，輸出 $50 — 旗艦級推理
GPT-6 Sol：輸入 $2，輸出 $10 — 比上一代Sol便宜50%
GPT-6 Luna：輸入 $0.10，輸出 $0.50 — 輸出成本比上一代砍58%

Luna的定位是高量、簡單任務的清潔工；Sol則主打「接近Astra但花不到兩成的錢」。

GPT-6 Sol真的打平Astra嗎？

官方與第三方測試（AutomationBench、OSWorld 2.0）顯示，GPT-6 Sol在「xhigh」努力設定下，桌面任務跑到60.5%，幾乎追平Claude Opus 5在中努力設定下的60.3%，但成本估計低了八成。整體來說，Sol能拿到Astra約90%-95%的實戰能力，價格卻只要兩成。

不過誠實講，這裡有個容易被行銷話術蓋過去的細節：要拿到接近Astra的分數，Sol必須開到xhigh努力等級，而不是預設值。Sol在xhigh下拿66.6%分要價$1.00，Astra在低努力下拿67%要價$1.60——差距其實沒有「五折」那麼夢幻，比較像八折到六折之間，實際省多少要看你把努力調到多高。

誰適合換GPT-6 Sol或Luna？

大量制式任務（客服分類、資料清洗、簡單摘要）：直接上Luna，成本砍到只剩零頭，品質對得起價錢。
中重度Agent任務（多步驟工具呼叫、程式碼審查）：Sol是划算選擇，但記得把努力設定調高，不要用預設值就拿它跟Astra比。
高風險、單次容錯率低的推理（法律、醫療、財務決策）：目前建議還是留在Astra，省下來的兩成成本，可能抵不過一次判斷失誤的代價。

OpenAI這次把GPT-6 Sol和Luna定成永久價而非促銷，代表它是想長期用價格戰卡位，而不是曇花一現的行銷手法。但「降價腰斬」跟「效能打平」是兩件事，選之前先看清楚自己要跑的任務用哪個努力等級，才不會表面省錢、實際上多花時間重試。

好不好用，試了才知道。

🇺🇸 GPT-6 Sol & Luna Review: Half Price, Same Power?

What Are GPT-6 Sol and Luna?

GPT-6 Sol and GPT-6 Luna are the two API models OpenAI shipped on September 22, 2026, trained the same way as flagship Astra but priced to compete on cost. OpenAI confirmed to press that the 50% price cut is a permanent rate, not a limited-time promo — the first time OpenAI has genuinely fought on price since Astra launched.

How Much Do GPT-6 Sol and Luna Cost?

The new three-tier lineup:

Astra: $10 input / $50 output per million tokens — flagship reasoning
GPT-6 Sol: $2 input / $10 output — 50% cheaper than the previous Sol
GPT-6 Luna: $0.10 input / $0.50 output — output cost down 58% from last gen

Luna is built for high-volume, low-complexity work. Sol's pitch is "near-Astra quality for a fifth of the price."

Does GPT-6 Sol Actually Match Astra?

Third-party benchmarks (AutomationBench, OSWorld 2.0) show GPT-6 Sol at "xhigh" effort hitting 60.5% on desktop tasks — nearly matching Claude Opus 5 at medium effort (60.3%) — at roughly 80% lower cost. Broadly, Sol delivers 90-95% of Astra's real-world capability at 20% of the price.

Here's the catch marketing tends to skip: hitting Astra-level scores requires running Sol at xhigh effort, not the default setting. Sol at xhigh scores 66.6% for $1.00 per task; Astra at low effort scores 67% for $1.60. That's closer to a 40% discount than a flat "half price," and your actual savings depend entirely on how high you dial the effort setting.

Who Should Actually Switch?

High-volume routine work (support triage, data cleanup, basic summarization): go straight to Luna — cost drops to pennies and quality holds up.
Multi-step agent tasks (tool-calling chains, code review): Sol is a solid deal, but crank effort up — comparing it to Astra at default settings isn't a fair test.
High-stakes, low-tolerance-for-error reasoning (legal, medical, financial decisions): stick with Astra for now — the 20% you save isn't worth one bad call.

Locking in the price cut permanently signals OpenAI wants to compete on cost long-term, not just grab a headline. But "half price" and "equal performance" are two separate claims — check which effort tier your workload actually needs before you migrate, or you'll save on tokens and lose it back in retries.

You won't know until you try it.

Sources / 資料來源

常見問題 FAQ

GPT-6 Sol和Luna什麼時候可以用？

已於2026年9月22日在ChatGPT Work、Codex與API同步上線，即刻可用，價格為永久定價。

GPT-6 Sol比GPT-6 Astra便宜多少？

輸入輸出token價格都只要Astra的兩成（$2/$10 vs $10/$50），但要達到接近Astra的效能需開到xhigh努力等級。

GPT-6 Luna適合拿來做什麼？

適合大量、簡單的制式任務，例如客服分類、資料清洗、基礎摘要，成本比前代砍58%輸出成本。

換成GPT-6 Sol會不會犧牲品質？

在xhigh努力設定下，Sol能拿到Astra約九成到九成五的實戰表現，但預設努力等級分數會明顯較低。

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