Use AI after 11 pm, free or half price
Smart Music launched a nightly use activity yesterday, and there is no deduction for GLM-5.3-Flash every night from 11:00 to 9:00 the next day. After it spread, many people thought it was "Smart Music is free at night." I flipped out the official announcement and read it again, and found that three prerequisites must be met at the same time: first, it must be paid for subscription users, free only in ZCode, and only GLM-5.3-Flash. What's more worth saying is another thing: this is not a promotion for one company. DeepSeek has long written half price for off-peak peaks on its pricing page, and the range of lows is wider. Every night, weekday lunch break, and the entire weekend are considered free. If you move the same task to run it at night, you can save half of it, and you can hit the cache and save 80% of it.
Yesterday, Smart Music launched an activity: from 11:00 every night to 9:00 a.m. the next day, there is no deduction for GLM-5.3-Flash.
After this matter spread, many people understood it as "wisdom music is free at night."
I went to read the official announcement and found that to be established, three conditions had to be met at the same time. What's more worth mentioning is another thing: this is not a promotion for one company, it's AI that started learning about electricity bills and sold them during the day and at night.
first made the three prerequisites clear
The original official announcement read as follows:
From September 3 to September 20, the GLM Coding Plan launched a nightly use activity. Every night from 23:00 to 9:00 the next day, you can call GLM-5.3-Flash in the plan and use it for free in ZCode. The amount of use in Agents supported by other plans will double, and it will automatically take effect during a specified time period,covering all paid plan users.
Three prerequisites are hidden in this passage.
First, you have to be a paying user first. The original text says "Covering all paid plan users", which means that this is the rights of Coding Plan subscribers, not free for everyone. If you don't subscribe, this event has nothing to do with you.
Second, free is only available in ZCode. The official form is clearly divided:
| use | quota consumption |
|---|---|
| Using ZCode | Limit consumption is all 0 |
| Other Agents supported through the package | The amount is ×2 times based on standard rules |
ZCode is their own programming software. If you use Claude Code or other Agent to connect in, you will get double the amount, not free. Doubling is nice, but it's two different things from "free".
Third, there is only one model, GLM-5.3-Flash. A special sentence was written in the announcement: Select GLM-5.3 during off-peak periods, and the consumption quota will still be based on the standard rules of the package. At the same time period, if you change the model, you will deduct it as usual.
The time is generous: from September 3rd to 20th, every night from 23:00 to 9:00 the next day,including weekends and public holidays, the entry period will automatically take effect, and there is no need to manually collect it.
Another family has done this for a long time, and it is for a long time.
Wisdom is a time-limited event that ends on September 20. But the same idea has been written on the pricing page by DeepSeek, which is a long-term mechanism.
Its original words were:
The > idle period price is half of the peak period price. Peak hours are 9:00 - 12:00 and 14:00 - 18:00 from Monday to Friday, Beijing time (the rest are idle periods)
Note that its peak definition is much narrower than the intellectual spectrum. Only the peak hours on weekdays are from 9 a.m. to 12 a.m. and 2 a.m. to 6 p.m., and the rest are all free: half price every night, lunch breaks on weekdays, and the entire weekend.
Specific to the price, take deepseek-v4-flash as an example, the unit is yuan per million tokens:
| peak hours | 高峰时段 | |
|---|---|---|
| Input (cache miss) | 1.5 | 3.0 |
| Input (cache hit) | 0.05 | 0.10 |
| output | 4.5 | 9.0 |
There is also a second price difference hidden here, which is even worse than a peak error: the input price for cache hits is 0.05, and the miss is 1.5, which is 30 times the difference.
Accounting
Suppose you want to run a task, enter 1 million tokens and output 200,000 tokens.
Peak period, no cache: 3.0 + 9.0×0.2 =4.8 yuan
Moved to the evening without caching: 1.5 + 4.5×0.2 =2.4 yuan
Move it to night and hit the cache: 0.05 + 0.9 =0.95 yuan
I didn't change anything, just moved the task from afternoon to evening,saving half of it. If your scene can still hit the cache,you will save 80% in total.
For individual users, this may be a matter of a few dollars. But if you are running batch tasks, such as sorting out a batch of documents, batch translation, and data cleaning, these tasks don't require you to keep an eye on them. If you run at night, the cost will be cut in half.
Why are both families rushing to night?
This logic is actually not new. The electricity bill has been calculated long ago.
Everyone is using it during the day, and the GPU cluster is tight; in the second half of the night, the machines are idle and idle. Using the price difference to drive part of the load into the night, the platform's cluster utilization rate increased, and the user's bills came down.
It is cost-effective for the platform: even if you give out the idle computing power at night for free, it is better than idle, and you can also pull users away from competitors.
For users, whether you can take advantage depends on whether your taskis urgent or not.
So how to use it
All the work that can be asynchronous can be moved to the night. Batch sorting, batch translation, data cleaning, and running test sets, these tasks that don't require anyone to keep an eye on them, just write a scheduled task and throw it to run after 23 o'clock. DeepSeek is priced at half price, and it is free all day on weekends.
For scenarios that can hit the cache, the price difference is greater than the peak error. If the same system prompt is used repeatedly and the same long document is asked multiple times, this kind of scenario has a high cache hit rate and the input price can be 30 times worse, which is more cost-effective than shifting time.
Only people who have already subscribed to a Coding Plan need to care about the Smart Spectrum activity. If you book, remember two things: using ZCode to tune Flash at night will cost zero consumption, and changing to other Agents will only double; the event ends on September 20.
If you haven't subscribed, don't be deceived by the words "free at night". It is the rights of subscribers, not a new free entrance.
few sentences boundary
The above rules all come from the official announcements and pricing pages of the two companies and are not second-hand paraphrases. Smart Spectrum is a time-limited event that ends on September 20;DeepSeek is written on the pricing page and is a long-term mechanism, but the price policy may be adjusted at any time.
I don't have a subscription to the Coding Plan, so I didn't personally verify the actual effect of zero consumption at night. What is written on it is the officially announced rules. DeepSeek's price is a public figure on the pricing page, and all I did was do the calculations.
I checked the arithmetic part: 4.8 divided by 2.4 is exactly twice, and 0.95 is 80% less than 4.8.