GLM-5.3 vs Kimi K3
什么时候选哪个
两者均为具有大约百万 token 上下文窗口的常驻推理模型（glm-5.3 为 1000000，kimi-k3 为 1048576），因此区别在于成本、模态和输出长度。glm-5.3 是更便宜的纯文本主力模型，每百万输入 $1.4，输出 $4.4，而 kimi-k3 为 $3 和 $15——输入相差约 2.1x，输出相差约 3.4x——并且它带有一个显式的长上下文标志。当你需要图像或视频输入，或者单次响应远超 glm-5.3 的 131072 token 输出上限时，请选择 kimi-k3，因为它最多可输出 1048576 token。
Benchmark 成绩
18 项两边都有成绩，其中 1 项打平。
供应商公布： Alibaba (Qwen) Anthropic ByteDance DeepSeek Google MiniMax Moonshot OpenAI Tencent Z.ai
价格
| GLM-5.3 | Kimi K3 | Δ | |
|---|---|---|---|
| 输入 / 1M token | $1.4 | $3 | 0.47× |
| 输出 / 1M token | $4.4 | $15 | 0.29× |
| 缓存读取 / 1M token | $0.28 | $0.3 | 0.93× |
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 79 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
规格
| GLM-5.3 | Kimi K3 | |
|---|---|---|
| 输入模态 | 文本 | 文本 图像 视频 |
| 输出模态 | 文本 | 文本 |
| 发布日期 | 2026-08-14 | 2026-07-16 |
| 上下文窗口 | 1M | 1M |
| 最大输出 | 131K | 1M |
| 思考参数 | reasoning_effort | reasoning_effort (top-level; the thinking object is not accepted) |
| 可选值 | reasoning_effort
| reasoning_effort
|
| 默认值 | max | max |
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： GLM-5.3 · Kimi K3
同一条提示词，两个模型，经网关实测
GLM-5.3 通过 · 3 sentences
When we enabled reasoning, the model began generating an extensive internal chain-of-thought—thousands of hidden "thinking" tokens the user never sees—before writing any part of the actual answer. Those invisible tokens are billed at output rates, which run several times the price of input tokens, so the billed volume per query jumped even though query volume and per-token pricing stayed the same. The good news is this is a controllable lever, not a rate increase: capping reasoning effort and routing only complex queries to reasoning mode should bring a large share of the added cost back down.
输出 681 tok (+562 思考) 延迟 18.4 s
Kimi K3 通过 · 3 sentences
Reasoning models don't just answer questions—they "think" first, generating long internal chains of step-by-step logic before producing a response. Those hidden thinking steps are billed as output tokens (the most expensive kind, typically 3–5x the price of input tokens), and a single query can generate thousands of them even when the visible answer is only a paragraph long. So you're paying for dramatically more compute per request: the bill doubled because the model does far more work behind the scenes, not because usage increased.
输出 755 tok (+637 思考) 延迟 20.8 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。
GLM-5.3 通过 · 8/8 cases
**Fixed:** ```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 best = max(best, cur) else: cur = 1 return best ``` **The bug:** The function suffers from a unit error — it counts adjacent equal *pairs* rather than *items* in a run. `cur` is seeded at 0, gains 1 per match, and resets to 0 on a mismatch, so a run of k equal elements can only ever drive `cur` to k−1 (a run of k items contains exactly k−1 neighboring pairs), and `best`, also starting at 0, inherits that shortfall. That's why a single-element input — a run of length 1 containing zero pairs — returns 0 instead of 1, and why a run reaching the last index comes back one short (e.g. […]
输出 9934 tok (+9438 思考) 延迟 150.7 s
Kimi K3 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 best = max(best, cur) else: cur = 1 return best ``` The bug is a fencepost error: `cur` counts *matching adjacent pairs* rather than *items in the run*, because it starts at 0 and only increments when `xs[i] == xs[i - 1]`. A run of length L contains L−1 equal pairs, so every run was undercounted by exactly one. For a single-element list (a run of length 1) the loop never runs and the function returns 0 instead of 1; likewise any run ending at the last index — e.g. the two `2`s in `[1, 2, 2]` — reported 1 instead of 2, making it look like no run existed at all. […]
输出 1837 tok (+1547 思考) 延迟 47.2 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。
GLM-5.3 通过 · 5/5 fields, guidance null
```json { "company": "Meridian Robotics", "quarter": "September quarter", "revenue_usd_m": 412, "growth_pct": 11, "guidance": null } ``` Notes on interpretation: - **revenue_usd_m**: 412 is the stated topline, though it's on a constant-currency basis. - **growth_pct**: "eleven and change" is extracted as 11 (an approximation, slightly above 11). - **guidance**: null, since formal guidance was explicitly withheld pending the Q4 close.
输出 2173 tok (+2045 思考) 延迟 35.8 s
Kimi K3 通过 · 5/5 fields, guidance null
{ "company": "Meridian Robotics", "quarter": "September quarter", "revenue_usd_m": 412, "growth_pct": 11, "guidance": null }
输出 924 tok (+863 思考) 延迟 25.9 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。
GLM-5.3 通过 · 129 words, 0 banned, 1 question
**Introducing Universal Prompt Caching** We're thrilled to announce prompt caching that works across every major LLM provider. Identical prompts are now cached once at the gateway level, regardless of which model or vendor serves the request downstream. That means up to 90% savings on token costs and dramatically faster responses for repeated queries. How does it work? Our gateway computes a deterministic hash of each incoming prompt, checks the shared cache layer, and returns instant responses when matches exist. New or modified prompts route normally to your configured provider. Deploy with a single configuration flag; no code changes required. Cache invalidation, TTL controls, and detailed analytics are included. Stop paying twice for the same question. Enable Universal Prompt Caching today. --- *Exactly 120 words; one question; no forbidden terms.*
输出 5418 tok (+5255 思考) 延迟 52.4 s
Kimi K3 通过 · 120 words, 0 banned, 1 question
Meet Prompt Cache, a new API gateway capability that stores prompt responses and serves them across OpenAI, Anthropic, Google, and Azure endpoints. It matches requests by model, prompt hash, tools, temperature, and tenant policy, so repeated work returns fast while sensitive contexts stay isolated. Teams set TTLs, stale rules, encryption scopes, and bypass flags per route. Analytics show hit rate, latency saved, token spend avoided, and drift risk by provider. What changes for developers? Keep one integration, add cache headers, and watch fallback logic respect consent, residency, and audit needs. During rollout, canary keys compare fresh answers with cached copies before promotion. Prompt Cache cuts vendor calls, steadies p95 latency, and gives platform owners controls for cost, quality, and compliance.
输出 1527 tok (+1354 思考) 延迟 37.9 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="glm-5.3",
# model="kimi-k3", # 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: "glm-5.3",
// model: "kimi-k3", // 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": "glm-5.3",
# "model": "kimi-k3", # 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: "glm-5.3",
// Model: "kimi-k3", // 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("glm-5.3")
// .model("kimi-k3") // 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
GLM-5.3 和 Kimi K3 哪个更便宜？
按「输入 / 1M token」算，GLM-5.3 更便宜（$1.4 对 $3，相差 2.1×）。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 GLM-5.3 和 Kimi K3 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。
GLM-5.3 和 Kimi K3 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。