GPT-5.6評測：Sol、Terra、Luna三檔怎麼選不踩雷 | GPT-5.6 Review: Sol vs Terra vs Luna, Pick Right

9月 15, 2026

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

🇹🇼 GPT-5.6評測：Sol、Terra、Luna三檔怎麼選不踩雷

OpenAI 最新發布的 GPT-5.6 大改了命名邏輯：不再是單一模型，而是切成 Sol、Terra、Luna 三個檔位，讓開發者依任務難度和預算自行搭配。三檔到底怎麼選、划不划算，直接影響 API 帳單，這篇用實測數字說清楚，不談行銷術語。

GPT-5.6 三檔是什麼

在 GPT-5.6 的新命名法裡，版本號代表世代，Sol/Terra/Luna 則是可各自獨立升級的能力層級：

Sol：旗艦款，唯一支援最高推理強度與 Ultra 模式，每百萬 token 輸入 5 美元、輸出 30 美元。
Terra：主力款，多數跑分只落後 Sol 2～3 分，價格砍半以上（輸入 2 美元／輸出 12 美元）。
Luna：極速款，價格只要 Sol 的二十五分之一（輸入 0.2 美元／輸出 1.2 美元），適合高流量、低深度的工作。

別只看單價，要算「解出一個任務的成本」

程式碼審查工具商 CodeRabbit 實測 100 多個長流程編碼任務發現：Sol 平均只用 20,968 個 output token 就拿下 63.7% 通過率；Terra 雖然單價便宜，卻要燒到 55,594 個 token 才拿到 40.7% 通過率。換算下來，複雜任務丟給 Terra 省的是帳面單價，賠的是總 token 量與失敗重跑的時間——便宜不等於划算。

三檔怎麼分工比較實際

Terra 當預設：一般任務、程式碼初審、中等複雜度的助理功能。
Sol 留給硬仗：長流程 Agent、瀏覽器與終端機操作（BrowseComp 拿下 92.2%、Terminal-Bench 2.1 拿下 88.8%）、資安研究這類需要一路盯到底的工作。
Luna 專攻走量：分類、摘要這種淺任務，但長文本記憶只有 41.3%（Sol 是 91.5%），別拿它做需要記住整份文件的工作。

誠實地說：GPT-5.6 不是全面最強

在純程式碼跑分 SWE-Bench Pro 上，Claude Fable 5 拿下 80%，Sol 只有 64.6%。GPT-5.6 的優勢在「長時間不掉線的執行力」，不是每項都拿第一。如果工作是策略規劃或架構判斷，Fable 5 或 Sonnet 5 可能更適合；如果是要一個 Agent 悶頭做完一長串任務不出包，Sol 才是對的選擇。

好不好用，試了才知道。

🇺🇸 GPT-5.6 Review: Sol vs Terra vs Luna, Pick Right

OpenAI just reshuffled how it ships models: GPT-5.6 isn't one model anymore, it's three tiers — Sol, Terra, and Luna — each aimed at a different budget and task depth. Picking the wrong one either wastes money or wastes an agent's follow-through. Here is what the numbers actually say, no marketing spin.

What Sol, Terra, and Luna Actually Are

In OpenAI's new scheme, the version number marks the generation; Sol/Terra/Luna are durable capability tiers that can each update on their own cadence:

Sol: the flagship, the only tier with max reasoning effort and Ultra mode. input / output per 1M tokens.
Terra: the workhorse, within 2-3 points of Sol on most benchmarks at less than half the price ( input / output).
Luna: the speed tier, roughly 1/25th of Sol's price (/bin/zsh.20 input / .20 output), built for high-volume, low-depth work.

Stop Comparing Unit Price — Compare Cost Per Solved Task

Code review vendor CodeRabbit ran 100+ long-horizon coding tasks and found Sol solved 63.7% of them using just 20,968 output tokens on average. Terra, despite its cheaper per-token rate, needed 55,594 tokens to land only 40.7%. On genuinely hard tasks, Terra's discount evaporates once you count total tokens burned and failed reruns — cheap and cost-effective aren't the same thing.

A More Practical Way to Split the Work

Default to Terra for everyday tasks, first-pass code review, and mid-complexity assistants.
Reserve Sol for long-horizon agents, browser/terminal operation (92.2% on BrowseComp, 88.8% on Terminal-Bench 2.1), and cybersecurity research — anything that needs to stay on track for a long stretch.
Use Luna only for shallow, high-volume work like classification and summarization. Its long-context recall is 41.3% versus Sol's 91.5%, so don't hand it anything that needs to remember a full document.

Honestly, GPT-5.6 Isn't the Best at Everything

On pure coding benchmarks like SWE-Bench Pro, Claude Fable 5 scores 80% against Sol's 64.6%. GPT-5.6's real edge is staying on task over long stretches, not topping every leaderboard. For architectural judgment or open-ended planning, Fable 5 or Sonnet 5 may still be the better call — Sol shines when you need an agent to grind through a long task list without falling apart.

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