GPT-5.5 vs GPT-5.6 Luna
什麼情況選哪一個
gpt-5.6-luna 便宜非常多：每百萬輸入 token $0.2 對 $5，輸出 $1.2 對 $30，快取讀取 $0.02 對 $0.5，每一項都便宜 25 倍；同時兩者的上下文都是 1050000 token，最大輸出都是 128000 token。gpt-5.5 保留了 computer use 與平行工具呼叫，而 gpt-5.6-luna 並未列出這兩項，所以做工具驅動或操作介面類的工作時，仍應固定使用 gpt-5.5；一般的文字與程式碼大量處理，則用新模型，在相同視窗下便宜 25 倍。
Benchmark 成績
兩邊都有成績的有 41 項。
供應商公布： Alibaba (Qwen) Anthropic ByteDance DeepSeek Google MiniMax Moonshot OpenAI Tencent Z.ai
定價
| GPT-5.5 | GPT-5.6 Luna | Δ | |
|---|---|---|---|
| 輸入 / 1M tokens | $5 | $0.2 | 25× |
| 輸出 / 1M tokens | $30 | $1.2 | 25× |
| 快取讀取 / 1M tokens | $0.5 | $0.02 | 25× |
| 快取寫入 | 不額外計費 | 不額外計費 | - |
費率取自網站建置時的即時目錄；各模型頁面都列有最新的價目。
兩者的相對位置：每 1M tokens 的輸入價格，涵蓋同一計費單位下全部 78 個聊天模型（對數尺度）
功能
| GPT-5.5 | GPT-5.6 Luna | |
|---|---|---|
| 工具使用 | 是 | 是 |
| 思考控制 | 可設定 | 可設定 |
| 結構化輸出 | 是 | 是 |
| 提示詞快取 | 隱式（自動） | 隱式（自動） |
| 快取存活時間 | 5-10m, up to 1h | 5-10m, up to 1h |
| 最小快取前綴 | 1024 個 token | 1024 個 token |
規格
| GPT-5.5 | GPT-5.6 Luna | |
|---|---|---|
| 輸入模態 | 文字 圖像 | 文字 圖像 |
| 輸出模態 | 文字 | 文字 |
| 發布日期 | 2026-04-24 | 2026-07-09 |
| 知識截止日期 | 2025-12 | 2026-02 |
| 上下文視窗 | 1.1M | 1.1M |
| 最大輸出 | 128K | 128K |
| 思考參數 | reasoning.effort | reasoning.effort |
| 可接受的值 | reasoning.effort
| reasoning.effort
|
| 預設值 | medium | medium |
規格摘錄自各供應商的文件；供應商沒有公布的項目就直接略過，不自行推測。 完整來源： GPT-5.5 · GPT-5.6 Luna
同一段提示詞，兩個模型，經閘道實測
GPT-5.5 通過 · 3 sentences
Enabling reasoning makes the model do additional internal “thinking” before producing the final answer, which consumes more compute and billable tokens. Even if user volume and output length stayed the same, each request now costs more because the model is spending extra capacity to solve the problem. In short, we bought higher-quality answers, but the unit cost per answer increased enough to double the total LLM bill.
輸出 131 tok (+42 思考) 延遲 2.9 s
GPT-5.6 Luna 通過 · 3 sentences
Enabling reasoning causes the model to use substantially more computational steps and generated tokens per request, even when the visible answer is similar. Because our provider charges based largely on token consumption and compute, that increased processing translated into roughly twice the cost. We can control the expense by reserving reasoning for complex tasks, setting token and effort limits, and monitoring cost per successful outcome.
輸出 146 tok (+60 思考) 延遲 3.1 s
是否照指令做（剛好三句，數得出來）、語氣是否貼合對象（寫給 CFO 看），以及下方 token 計量顯示出的隱藏思考計費落差。
GPT-5.5 通過 · 8/8 cases
```python def longest_run(xs): if not xs: return 0 best = 1 cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 else: cur = 1 best = max(best, cur) return best ``` The bug was that `cur` and `best` were counting the number of equal *adjacent comparisons*, not the number of items in the run. A run of one item has length `1`, but the original code initialized both values to `0`, so single-element inputs returned `0`. Similarly, a run like `[2, 2, 2]` has length `3`, but only two adjacent equal comparisons, so the result was off by one. Also, `best` should be updated after every element is processed, not only inside the equality branch, so runs that continue through the final index are properly considered.
輸出 401 tok (+182 思考) 延遲 5.9 s
GPT-5.6 Luna 通過 · 8/8 cases
```python def longest_run(xs): if not xs: return 0 best = cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 else: best = max(best, cur) cur = 1 return max(best, cur) ``` The bug is that `cur` counts equal *transitions* rather than the number of items in the current run, so a single-element run returns `0` and a run of length `n` returns `n - 1`; additionally, `best` is only updated when a run is interrupted, so a run continuing through the final element is not finalized before returning. Initializing the current run to one item, resetting it to one after a change, and taking a final `max` fixes both issues.
輸出 356 tok (+161 思考) 延遲 4.7 s
修正是否真的正確（跑得起來）、說明的資訊密度，以及在範圍明確的任務上的 token 效率。
GPT-5.5 通過 · 5/5 fields, guidance "withheld pending the Q4 close"
{"company":"Meridian Robotics","quarter":"September quarter","revenue_usd_m":412,"growth_pct":11,"guidance":"withheld pending the Q4 close"}
輸出 557 tok (+512 思考) 延遲 6.7 s
GPT-5.6 Luna 通過 · 5/5 fields, guidance "Withheld pending the Q4 close"
{"company":"Meridian Robotics","quarter":"September quarter","revenue_usd_m":412,"growth_pct":null,"guidance":"Withheld pending the Q4 close"}
輸出 227 tok (+181 思考) 延遲 3.6 s
是否遵守 schema（不自創欄位）、面對幻覺誘因的表現（題目明確不給 guidance），以及結構化輸出路徑的差異。
GPT-5.5 通過 · 120 words, 0 banned, 1 question
Today we’re introducing Prompt Cache Routing, a new API gateway feature that stores reusable prompt prefixes and applies them across leading model providers. Teams can reduce latency, lower token spend, and keep application logic portable without rewriting requests for each vendor. Have you ever paid twice to send the same system instructions? With shared cache keys, provider-aware metadata, and automatic fallback handling, the gateway detects eligible prompt segments, reuses cached context, and records savings in your existing observability dashboards. Policies let admins set retention windows, data boundaries, and provider allowlists by workspace or environment. Prompt Cache Routing is available now in beta for Pro and Enterprise customers, with SDK support, Terraform resources, and clear migration guides included at launch worldwide.
輸出 859 tok (+700 思考) 延遲 8.7 s
GPT-5.6 Luna 通過 · 120 words, 0 banned, 1 question
Introducing PromptCache, an API gateway feature that caches prompts across providers, helping teams reduce latency, control spend, and deliver consistent results. How much faster could your applications respond when repeated prompts are served from a shared cache instead of being sent upstream? PromptCache supports provider-aware routing, configurable time-to-live policies, encrypted storage, cache invalidation, and usage analytics through one operational layer. It works with language-model providers while preserving your authentication, observability, and fallback workflows. Developers can enable caching by endpoint, model, tenant, or prompt pattern, then monitor hit rates and savings in real time. Built for production workloads, PromptCache gives platform teams controls for performance and cost without requiring application rewrites. […]
輸出 948 tok (+778 思考) 延遲 8.0 s
是否遵守限制（字數上限、禁用詞清單、只能有一句問句）、文字風格的特徵，以及長度控制。
改一行程式碼就能在兩者之間切換
下面每個頁籤都列了這兩個模型 ID，要改的只有醒目標示的那兩行。端點、金鑰和請求格式都不變。
from openai import OpenAI
client = OpenAI(
base_url="https://synthorai.io/v1",
api_key="sk-syn-...",
)
resp = client.chat.completions.create(
model="gpt-5.5",
# model="gpt-5.6-luna", # 取消這一行的註解，並把上一行註解掉
messages=[{"role": "user", "content": "Summarize this diff"}],
reasoning_effort="medium",
)
print(resp.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://synthorai.io/v1",
apiKey: "sk-syn-...",
});
const resp = await client.chat.completions.create({
model: "gpt-5.5",
// model: "gpt-5.6-luna", // 取消這一行的註解，並把上一行註解掉
messages: [{ role: "user", content: "Summarize this diff" }],
reasoning_effort: "medium",
});
console.log(resp.choices[0].message.content);curl https://synthorai.io/v1/chat/completions \
-H "Authorization: Bearer sk-syn-..." \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-5.5",
# "model": "gpt-5.6-luna", # 取消這一行的註解，並把上一行註解掉
"messages": [{"role": "user", "content": "Hello"}],
"reasoning_effort": "medium"
}'package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/option"
)
func main() {
client := openai.NewClient(
option.WithBaseURL("https://synthorai.io/v1"),
option.WithAPIKey("sk-syn-..."),
)
resp, _ := client.Chat.Completions.New(context.TODO(), openai.ChatCompletionNewParams{
Model: "gpt-5.5",
// Model: "gpt-5.6-luna", // 取消這一行的註解，並把上一行註解掉
Messages: []openai.ChatCompletionMessageParamUnion{
openai.UserMessage("Summarize this diff"),
},
ReasoningEffort: openai.ReasoningEffortMedium,
})
fmt.Println(resp.Choices[0].Message.Content)
}import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.chat.completions.*;
import com.openai.models.ReasoningEffort;
OpenAIClient client = OpenAIOkHttpClient.builder()
.baseUrl("https://synthorai.io/v1")
.apiKey("sk-syn-...")
.build();
ChatCompletion resp = client.chat().completions().create(
ChatCompletionCreateParams.builder()
.model("gpt-5.5")
// .model("gpt-5.6-luna") // 取消這一行的註解，並把上一行註解掉
.addUserMessage("Summarize this diff")
.reasoningEffort(ReasoningEffort.MEDIUM)
.build());
System.out.println(resp.choices().get(0).message().content().orElse(""));常見問題
GPT-5.5 和 GPT-5.6 Luna 哪個比較便宜？
以「輸入 / 1M tokens」來看，GPT-5.6 Luna 比較便宜（$0.2 對 $5，相差 25×）。其他項目的結果可能相反，完整價目請看上表；實際成本要看你的用量組合。
可以只串接一次，就對 GPT-5.5 和 GPT-5.6 Luna 做 A/B 測試嗎？
可以。兩個模型都走同一個 OpenAI 相容端點，用的也是同一把 API 金鑰，切換時只要改一行裡的模型名稱字串。你可以把一部分流量分別導到兩邊，再直接比較帳單。
GPT-5.5 與 GPT-5.6 Luna 支援提示詞快取嗎？
支援。兩者的快取讀取費率都低於輸入費率，所以前綴已經進快取的工作負載，實際成本會比官網價算出來的低。確切的快取讀取價格請見上方定價表。