Can I replace the $20 ChatGPT Plus for free?

ChatGPT Plus costs $20 a month, which is about the same price as unlocking the latest GPT-5.6 Sol;Claude Pro, and unlocking the Opus 5, which was just released at the end of July. More than two hundred dollars a year. Can a free large model save you this money?

I don't want to just say "can" or "can't" to get it done. I took a few tough questions, searched for the top open source models that could be obtained for free, and put out their answers without changing a word. You can see for yourself whether they are enough to replace paying.

Let me be clear: I really tune Kimi K3 and DeepSeek-V4-Pro, which are the strongest open source models you can get for free now. For the reference on the payment side, I use myself (Claude Opus 4.8). To be honest, it is not the latest Opus 5 yet, but it is also a strong model that costs money and can be an anchor point. The two latest flagships, GPT-5.6 and Opus 5, require real money subscriptions. I can't adjust them directly, so I won't test them hard. I only use them as background when calculating "what I bought for 20 dollars".

The following is the process of true tuning, and I will post their original words for each question.

Path 1: Price trap

Title: The price of a product first increases by 20%, and then reduces the price by 20%. What is the final price of the original price?

Where is this question? Many people will blurt out,"It remains unchanged. If we increase by 20% and decrease by 20%, we will go back." Wrong. After a 20% increase, the base becomes larger, and a further 20% decrease is the larger number, so it is ultimately lower than the original price. The correct answer is 96%.

Kimi K3's answer: "Set the original price at 100. Price increase by 20%: 100 × 1.2 = 120. Reduce the price by another 20%: 120 × 0.8 = 96. So the final price was 96% of the original price, which was 4% lower than the original price."

DeepSeek-V4-Pro is more straightforward, directly giving "the final price is 96% of the original price." The process is also correct.

They solved both questions steadily, with clear formulas and correct conclusions. At this level, the top paid level is the same, and there is no gap.

Second Path: In-house sophistry "The Missing 1 Yuan"

This is an old-fashioned round-about-people question, but the question is a bit long: three people stay in the store, each person pays 10 yuan for a total of 30 yuan. The boss said that he would refund 5 yuan at a discount and asked the waiter to refund it. The waiter secretly hid 2 yuan and only refunded 1 yuan per person. So each person actually paid 9 yuan, and the total for the three people was 27 yuan. Including the 2 yuan hidden by the waiter, it was 29 yuan. But the original payment was 30 yuan. Where did the remaining 1 yuan go?

This question has confused countless people. Its trick is to add up the "27 yuan actually paid by the three people" and the "2 yuan hidden by the waiter" to get 29, and then compare it to 30, creating the illusion of "1 yuan less." But this addition is meaningless at all. The 27 yuan already includes the 2 yuan hidden by the waiter.

See how the free model is disassembled.

Kimi K3's answer: "'Money paid'(27) and 'Money received by the waiter'(2) are the relationship between the whole and part and cannot be directly added. The correct identity is always: total paid = total received by each party, which is 27 = 25 (room fee)+ 2 (waiter), or 30 = 25 + 2 + 3 (refund). The routine for this kind of question is to steal the concept and use a seemingly reasonable but meaningless formula to induce you to suspect that you have lost your money."

Not only was it correct, but it also exposed the mechanism of this question in one sentence: stealing the concept. This is true understanding, not just nonsense.

DeepSeek-V4-Pro's perspective is slightly different: "The 2 yuan hidden by the waiter is part of the 27 yuan. You cannot use the 27 yuan to add 2 yuan, but you should use the 27 yuan to add the returned 3 yuan, that is, 27 + 3 = 30 yuan. The 'missing one yuan' is entirely due to the illusion created by using the wrong formula."

The two free models, one divided from "whole and parts" and the other divided from "how to add correctly", have the same goal, and both neatly separated this sophistry question. This depth of understanding is already at the top level.

Third Route: Sally Sisters, by the way, take a look at free and classified

Title: Sally has 3 brothers, and each brother has 2 sisters. How many sisters does Sally have?

The correct solution is 1. Key point: Three brothers share the same group of sisters."Each brother has 2 sisters" means that there are only 2 girls in the family. Sally is one of them, so she only has 1 sister.

Kimi K3's answer: "1. Sally is a girl, not her sister. 2. Each brother has two sisters, which means that there are two girls in this family. 3. One of the two girls is Sally herself. 4. So in addition to Sally, there is still one girl, Sally's sister. Verification: There are 3 brothers + 2 girls in the family, and each brother sees these 2 sisters, which conforms to the meaning of the question. Final number: 1."

The reasoning chain is complete and verified by itself. Beautiful.

Why should I talk about this specifically? Because it exposes one thing by the way: free large models are also classified, don't go all the way. It is also free, but the GLM-4-Flash, which is fast and cheap, and which I recommended in the previous articles as the main force in collecting wool, has turned over. Its original words were: "Every brother has a sister who is Sally, and Sally has three brothers, so Sally has at least three sisters." Add Sally herself, 3 + 1 = 4." I went around my own logic and answered a 4.

There is also a similar machine problem (5 machines make 5 parts in 5 minutes, how many minutes does it take 100 machines make 100 parts, and 5 minutes to correct the problem). GLM-4-Flash took 1 minute to answer, and even Gemini took 100 minutes to answer. Kimi K3 and DeepSeek got it right for 5 minutes.

Therefore,"whether free can replace payment" must be made clear: what can be replaced is the top among free (Kimi K3, DeepSeek-V4), not all free ones. Fast and cheap lightweight versions (GLM-4-Flash, Gemini Flash-Lite), you can't pass the reasoning barrier.

!

capability level, words can be said to death

Look at the three types of problems: price hikes and price reductions trap, in-house sophistry, and reasoning questions. The free top-notch Kimi K3 and DeepSeek-V4-Pro all answered correctly. The reasoning process is clear and the mechanism of the problem can be pointed out. The few steps I did with Claude Opus 4.8 are at this level.

The conclusion is very real: in these daily tasks, free top jobs do as well as paid top jobs. If you spend 20 dollars, what you buy is no longer a "smarter brain." Free brains are enough.

What exactly does the 20 yuan buy?

I bought those free "tempers" that I really tested in the previous few articles, and paid to help you get rid of them.

Free ones need to limit current. There was a measured measurement in the article Collecting Wool. There were also 15 requests in a row. GLM-4-Flash accepted all 15 requests and collected them casually. However, GLM-4.7-Flash received 429 requests in a row, and fourteen out of fifteen requests were rejected; What's even more outrageous is Cerebras. It sounds generous to give you one million tokens a day, but in fact, the second request was restricted to death, and only fifteen requests were made into three. If you want to use it in batches or in series, you will be given a free look at any time. Open to pay.

Free and occasionally make it up seriously. In the last article, I wrote a non-existent Python library hyperfetch and asked them. GLM-4-Flash, Groq, Cerebras, and Gemini are fast and free, all of which have compiled functions and codes, and none of them said "there is no such library." Only Kimi K3, who can reason, was alerted while thinking,"I don't know about this library, it may not exist." Paid top-level models, especially the latest Opus 5 and GPT-5.6, are much more stable in this regard.

Free slowness. An inference model like DeepSeek, which thinks first and then answers, takes more than 20 seconds to think about a question. Pay quickly and strongly.

Also, the free window is closed at any time. If you can collect money today, you may be limited tomorrow; what you pay for is official guarantees, which are stable, up-to-date, multi-modal, and have a longer context.

!

Should these 20 yuan be saved

See what you lack.

Your daily life is mainly about answering and answering, writing something, translating, checking information, and writing simple code. You can accept it slower, check key information yourself, and you are willing to prepare Kimi K3 and DeepSeek for two free top-notch exchanges. That's enough for free to replace ChatGPT Plus, saving two to three hundred dollars a year, real money.

What you want is speed, stability, the latest and strongest brain (Opus 5 and GPT-5.6 are indeed one notch better than the free ones), multi-modal, and no need to worry about trouble. The 20-dollar purchase is exactly these, worth it.

In the final analysis, the price paid for 20 dollars is no longer "smart". The top free ones are smart enough to never fail a single problem. It sells "peace of mind" and "top matching". If you lack intelligence, you can manage it for free; if you don't want it, you can pay for it. Don't carry a lot of current limits and nonsense just to save money, and don't waste money for those rare top matches. Follow your own work.