GPT-5.4 vs GPT-5.6 Luna
什么时候选哪个
从这些信息看，两者只在价格和日期上有差别：都支持文本和图像输入，上下文窗口都是 1050000 token，最大输出都是 128000 token，也都可以关闭思考模式。gpt-5.6-luna 的每一项价格都更低：每百万 token 输入 $0.2、输出 $1.2、缓存输入 $0.02，而 gpt-5.4 为 $2.5、$15 和 $1.25，输入和输出价格是 gpt-5.6-luna 的 12.5 倍，缓存读取价格是 62.5 倍。优先选 gpt-5.6-luna，它的知识截止日期也更晚，为 2026-02（gpt-5.4 为 2025-08）；只有想继续使用较早的版本时才选 gpt-5.4。
Benchmark 成绩
14 项两边都有成绩。
供应商公布： Alibaba (Qwen) Anthropic ByteDance DeepSeek Google MiniMax Moonshot OpenAI Tencent Z.ai
价格
| GPT-5.4 | GPT-5.6 Luna | Δ | |
|---|---|---|---|
| 输入 / 1M token | $2.5 | $0.2 | 13× |
| 输出 / 1M token | $15 | $1.2 | 13× |
| 缓存读取 / 1M token | $1.25 | $0.02 | 63× |
| 缓存写入 | 不单独收费 | 不单独收费 | - |
价格取自构建时的实时目录，最新价格见各模型页面。
两个模型所处的位置：全部 78 个按同一单位计费的聊天模型的每 1M token 输入价分布（对数刻度）
能力
| GPT-5.4 | GPT-5.6 Luna | |
|---|---|---|
| 工具调用 | 是 | 是 |
| 思考控制 | 可配置 | 可配置 |
| 结构化输出 | 是 | 是 |
| 提示词缓存 | 隐式（自动） | 隐式（自动） |
| 缓存有效期 | 5-10m, up to 1h | 5-10m, up to 1h |
| 最小缓存前缀 | 1024 个 token | 1024 个 token |
规格
| GPT-5.4 | GPT-5.6 Luna | |
|---|---|---|
| 输入模态 | 文本 图像 | 文本 图像 |
| 输出模态 | 文本 | 文本 |
| 发布日期 | 2026-03-05 | 2026-07-09 |
| 知识截止日期 | 2025-08 | 2026-02 |
| 上下文窗口 | 1.1M | 1.1M |
| 最大输出 | 128K | 128K |
| 思考参数 | reasoning.effort | reasoning.effort |
| 可选值 | reasoning.effort
| reasoning.effort
|
| 默认值 | none | medium |
规格照录自各供应商的文档；供应商没有公布的项目，对应的行直接省略，不做推断。 完整来源： GPT-5.4 · GPT-5.6 Luna
同一条提示词，两个模型，经网关实测
GPT-5.4 通过 · 3 sentences
Enabling reasoning makes the model generate and evaluate additional internal steps before producing an answer, which increases the total tokens processed and therefore the cost per request. It also tends to increase latency and can trigger more tool calls, retries, or longer context retention, all of which compound usage-based charges. So even if user traffic stayed flat, the average cost per interaction rose materially once reasoning was turned on.
输出 82 tok 延迟 1.9 s
GPT-5.6 Luna 通过 · 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.
输出 146 tok (+60 思考) 延迟 3.1 s
指令遵循（恰好三句，数得出来）、受众适配（对 CFO 说话的口吻），以及下方 token 计数暴露出的隐藏思考计费差额。
GPT-5.4 通过 · 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: best = max(best, cur) cur = 1 best = max(best, cur) return best ``` The bug is that the original code was counting *matching transitions* between adjacent elements instead of the *length of the run itself*, so a single element started at `0` instead of `1`, making all run lengths off by one. It also only updated `best` inside the equality case, which means the final run could be lost if the loop ended before a change triggered any correction logic. […]
输出 243 tok 延迟 3.0 s
GPT-5.6 Luna 通过 · 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.
输出 356 tok (+161 思考) 延迟 4.7 s
修复是否真的正确（能运行）、解释的信息密度，以及在一个范围明确的任务上的 token 效率。
GPT-5.4 通过 · 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"}
输出 39 tok 延迟 1.3 s
GPT-5.6 Luna 通过 · 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"}
输出 227 tok (+181 思考) 延迟 3.6 s
是否严守 schema（不臆造字段）、能否顶住幻觉压力（原文明说暂不给出 guidance），以及结构化输出路径的差异。
GPT-5.4 通过 · 126 words, 0 banned, 1 question
Today we’re introducing Prompt Cache, a new API gateway feature that stores prompt results and reuses them across supported model providers. Teams can reduce latency, lower repeated compute costs, and keep application behavior more consistent when identical or near-identical prompts appear in production. Prompt Cache works at the gateway layer, so developers can adopt it without rewriting provider integrations or changing routing logic. Built-in controls let you set cache keys, expiration policies, and privacy rules for sensitive workloads. Need faster responses without duplicating optimization work for each vendor? With shared observability, you can track hit rates, savings, and cache performance in one place. […]
输出 154 tok 延迟 2.9 s
GPT-5.6 Luna 通过 · 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. […]
输出 948 tok (+778 思考) 延迟 8.0 s
是否守住约束（字数预算、禁用词表、只能有一个问句）、文风特征，以及长度控制。
改一行代码就能在两个模型之间切换
下方每个标签页里都有两个模型 ID，高亮的那两行是唯一要改的地方。端点不变，API key 不变，请求结构也不变。
from openai import OpenAI
client = OpenAI(
base_url="https://synthorai.io/v1",
api_key="sk-syn-...",
)
resp = client.chat.completions.create(
model="gpt-5.4",
# model="gpt-5.6-luna", # 取消注释此行，注释上一行
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-5.4",
// model: "gpt-5.6-luna", // 取消注释此行，注释上一行
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-5.4",
# "model": "gpt-5.6-luna", # 取消注释此行，注释上一行
"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-5.4",
// Model: "gpt-5.6-luna", // 取消注释此行，注释上一行
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-5.4")
// .model("gpt-5.6-luna") // 取消注释此行，注释上一行
.addUserMessage("Summarize this diff")
.reasoningEffort(ReasoningEffort.MEDIUM)
.build());
System.out.println(resp.choices().get(0).message().content().orElse(""));常见问题
GPT-5.4 和 GPT-5.6 Luna 哪个更便宜？
按「输入 / 1M token」算，GPT-5.6 Luna 更便宜（$0.2 对 $2.5，相差 13×）。其他计费项的结论可能相反，完整价格见上方表格，实际成本取决于你的用量构成。
不用分别集成两次，就能对 GPT-5.4 和 GPT-5.6 Luna 做 A/B 测试吗？
可以。两个模型走同一个 OpenAI 兼容端点，用同一个 API key，切换时只要改一行里的模型名，所以可以给两个模型各分一部分流量，直接对比账单。
GPT-5.4 和 GPT-5.6 Luna 支持提示词缓存吗？
支持。两个模型的缓存读取价都低于各自的输入价，所以前缀能反复命中缓存的负载，实际成本会比按官网价估算的低。具体的缓存读取价见上方价格表。