本文来源:https://mp.weixin.qq.com/s/u3nqWLFI4nYqG58q6UcfHA
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
官网充 28 块钱百万 token,有人用 1.9 块拿下了同款模型
写代码的人这两年应该都碰过这个选择题:选哪个模型?
选 DeepSeek 吧,便宜是便宜,但写复杂项目有时候差点意思。选 GPT 吧,好用是好用,价格摆在那,一个月下来账单够吃几顿好的。选智谱的 GLM 吧,能力确实强,但一看价格又缩回去了。
GLM-5.2 最近很火。
智谱最新的旗舰模型,1M 上下文,号称能吞下整个项目代码库,一次任务从需求到部署全干完。代码能力拉满了。去官网看了一眼价格——输入 8 块/百万 token,输出 28 块/百万 token。
写一篇文章可能就几毛钱,但要是天天用、大量用,一个月下来不是小数目。
前几天发现一个事:有人在云知声的平台上用 1.9 块就拿下了 GLM-5.2,还送了一堆其他模型。
● 平台叫 Token Hub
云知声做的 MaaS 平台,说实话以前对云知声的印象还停留在语音识别那块的,没想到他们搞了个大模型聚合平台。支持的模型还挺全:
- 文本旗舰:U2(云知声自研)、GLM-5.2(智谱)、Kimi K3(即将上线)
- 医疗模型:U2-Med、U1-OCR-Med
- 语音模型:ASR、TTS、声音克隆
- 视觉模型:OCR
关键是这些模型全在一个套餐里,不用每个单独付费。
● 算一笔账
先看智谱官方的价格:
|
|
|
|
|---|---|---|
|
|
|
28 |
|
|
|
|
|
|
|
|
GLM-5.2 输出 28 块/百万 token,写一篇几千字的文章,输出大概几千 token,算下来几毛钱一篇。看着不多,但如果每天都用,一个月干下来几十块到上百块,也不算少。
再看 Token Hub 的套餐:
|
|
|
|
|
|---|---|---|---|
| Lite | 1.9 元/月 |
|
1.8 亿 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1.9 块一个月,1.8 亿 Credits,GLM-5.2、U2、语音、OCR 全包。
对比一下就懂了。在智谱官网充 28 块只够买 100 万输出 token。在 Token Hub 花 1.9 块,拿到的是一整个月的套餐,里面装的模型还不止一个。
当然计费逻辑不一样。官网是按量计费用多少扣多少,这里是套餐制 Credits 共享。但 1.9 块这个数字放在那,怎么看都是碾压级的。
● 实测了下 GLM-5.2 和 U2
云知声还把自家的 U2 和 GLM-5.2 做了个组合:U2 管 Agent 能力,GLM-5.2 管代码能力。
U2 的 benchmark 数据摆在那:
|
|
|
|
|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
U2 的参数不大——稀疏 MoE 架构,266 亿总参只激活 10B——但几个关键维度都跑赢了同级别的竞品。
GLM-5.2 就更不用说了,智谱最新的代码旗舰,1M 上下文,项目级别的代码工程能力。这俩放一起,一个管 Agent 调度,一个管代码生成,日常开发和写作基本全覆盖了。
● 怎么用
流程很简单,三步:
第一步,去 maas.unisound.com 注册,订 Lite 套餐,1.9 块。
第二步,控制台拿 API Key 和 Base URL。Base URL 是 https://maas-api.unisound.com/v1,兼容 OpenAI 格式。
第三步,配到常用的工具里。Claude Code、OpenCode、Cursor、或者自己写的脚本,直接替换 Base URL 就行。
如果想在 cc-switch 这类路由工具里配,模型 ID 填 u2 或者 glm-5.2,跟 OpenAI 格式一样的接法。
● 说几句实话
1.9 块是首购价,次月续费是 19 块。19 块一个月 1.8 亿 Credits,对于日常写代码、写文章来说,正常使用根本用不完。
说是 1.9 块拿下 GLM-5.2,更准确的说法是 1.9 块拿下了整个平台的访问权,GLM-5.2 只是里面其中一个模型。
还有一个注意点:Credits 换算规则对不同模型可能不一样,订阅之前看一眼详情页的说明比较好。
但话说回来,1.9 块买不了吃亏买不了上当。一杯蜜雪冰城的钱,换个平台试一个月,GLM-5.2、U2 全都能用,不香吗?
● 附:各模型价格参考(按量计费)
怕有人需要按量计费的方案,也列一下:
|
|
|
|
|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
按量计费适合用量不稳定、怕绑死套餐的人。但要是确定每个月都用,套餐比按量划算太多。
This article is from: https://mp.weixin.qq.com/s/u3nqWLFI4nYqG58q6UcfHA If there is a copyright issue, please notify: 13951431913#139.com or submit a comment on the right.
Charge 28 yuan for a million tokens on the official website, someone got the same model for 1.9 yuan
People who write code should have encountered this choice in the past two years: which model to choose?
Choose DeepSeek, it is cheap, but for complex projects, it is sometimes lacking. Choose GPT, it is good, but the price is there, the bill for a month is enough for several good meals. Choose GLM from Zhipu, the ability is indeed strong, but looking at the price, it shrinks back.
GLM-5.2 is very popular recently.
The flagship model of Zhipu, with a 1M context, claims to be able to swallow the entire project code base and finish the whole task from requirement to deployment. The code ability is maxed out. I went to the official website and checked the price — input 8 yuan/million tokens, output 28 yuan/million tokens.
Writing an article may only cost a few cents, but if you use it every day and in large quantities, the total for a month is not a small amount.
I found out something recently: someone got GLM-5.2 on Unisound’s platform for 1.9 yuan, and they got a bunch of other models as well.
● The platform is called Token Hub
Made by Unisound, to be honest, my impression of Unisound was still in the field of speech recognition, but now they have made an aggregation platform for large models. The models they support are quite complete:
- Text flagship: U2 (Unisound’s own), GLM-5.2 (Zhipu), Kimi K3 (coming soon) - Medical models: U2-Med, U1-OCR-Med - Speech models: ASR, TTS, voice cloning - Vision models: OCR
The key is that all these models are in one package, no need to pay for each one separately.
● Do the math
First look at the official price of Zhipu:
Model
Input (RMB/million tokens)
Output (RMB/million tokens)
GLM-5.2
8
GLM-5.1
6 ~ 8
24 ~ 28
GLM-4.7
2 ~ 4
8 ~ 16
GLM-5.2 outputs 28 yuan/million tokens, writing an article of a few thousand words, output about a few thousand tokens, that would be a few cents per article. It does not look like much, but if you use it every day, it adds up to dozens or hundreds of yuan in a month.
Now look at the package of Token Hub:
Package
First purchase price
Renewal price
Credits
19 yuan/month
Standard
34.3 yuan/month
49 yuan/month
600 million
Pro
111.3 yuan/month
159 yuan/month
2.1 billion
Max
230.3 yuan/month
329 yuan/month
4.8 billion
For 1.9 yuan a month, 180 million credits, GLM-5.2, U2, speech, OCR are all included.
Compare them and you will understand. On the official Zhipu website, charging 28 yuan only buys 1 million output tokens. On Token Hub, spending 1.9 yuan gets you a whole month’s package, and there are more models inside than just one.
Of course, the billing logic is different. The official one is pay-as-you-go, deduct based on how much you use. Here it is a package system sharing credits. But looking at the number 1.9, it is a crushing victory.
● Tested GLM-5.2 and U2
Unisound also combined their own U2 and GLM-5.2: U2 handles agent ability, GLM-5.2 handles code ability.
The benchmark data for U2 is right there:
Dimension
U2 score
Comparison
Long context (AA-LCR)
70
> GLM-5.1 62.3
Knowledge reasoning (GPQA)
87.9
> GLM-5.1 86.2
Instruction following (IFBENCH)
77.3
> GLM-5.1 76.3
Tool calling (Claw-Eval)
76.9
> DeepSeek-V4 74
U2’s parameters are not large — sparse MoE architecture, 26.6 billion total parameters, only activating 10B — but it outperforms the same-level competitors in several key dimensions.
GLM-5.2 is even more needless to say, the latest code flagship of Zhipu, 1M context, project-level code engineering ability. Putting these two together, one handles agent scheduling, one handles code generation, basically covers all daily development and writing needs.
● How to use
The process is very simple, three steps:
First step , go to maas.unisound.com to register, order the Lite package, 1.9 yuan.
Second step , get the API Key and Base URL from the console. The Base URL is https://maas-api.unisound.com/v1 , compatible with OpenAI format.
Third step , configure it into the tools you use. Claude Code, OpenCode, Cursor, or your own scripts, directly replace the Base URL.
If you want to configure it in routing tools like cc-switch, fill in the model ID as u2 or glm-5.2 , same connection method as OpenAI format.
● A few honest words
The price of 1.9 yuan is the first purchase price , the renewal price for the next month is 19 yuan. 19 yuan a month for 180 million credits, for normal use of coding and writing articles, you will probably not even use it all up.
Saying you got GLM-5.2 for 1.9 yuan is actually saying you got access to the entire platform for 1.9 yuan. GLM-5.2 is just one model among them.
There is also one point to note: the conversion rules for credits may be different for different models, it is best to check the details page for explanations before subscribing.
But on second thought, for 1.9 yuan, you can’t lose. The price of a cup of coffee, try a platform for a month, GLM-5.2, U2 are all available, why not?
● Appendix: Reference prices for each model (pay-as-you-go)
For those who need pay-as-you-go plans, here they are:
Model
Input
Output
U2
1 yuan/million tokens
2 yuan/million tokens
U2-Med
2.5 yuan/million tokens
10 yuan/million tokens
U1-OCR
—
0.9 yuan/million tokens
U2-ASR
0.4 yuan/hour
—
U2-TTS
0.4 yuan/10,000 characters
—
Pay-as-you-go is suitable for people with unstable usage, who are afraid of being tied to a package. But if you are sure to use it every month, the package is much more cost-effective than pay-as-you-go.