Reviewing My AI Subscriptions

Machines never tire and never complain. —Kevin Kelly, What Technology Wants

---

In one day, I subscribed to two AIs back to back. Looking back, I realized I'd quietly ended up subscribing to this many AI models.

Claude

Claude is my favorite model right now. I'm on the Pro tier, 20 US dollars a month.

It has few flaws. Aside from Anthropic's international stance and its pricing, Claude is about as perfect as an AI model gets.

• In chat, Claude's conversation style is different from Gemini's and ChatGPT's. It's more serious, and more often than not it isn't a flatterer but a coach. Talking with it often produces new insights.

• Its Chinese is very good, and most of the time it's concise, with few holes in its logic.

• In Claude Code, who can doubt it's the best coding product in the universe right now? With Claude Code I built my blog from scratch and also developed a fairly mature product.

• On taste, Claude is excellent. Even if GPT's coding ability is on par with Claude's, on taste the two are worlds apart.

Put simply, Claude stands out partly because of its raw ability and partly because of its taste.

So if I could recommend only one AI model to anyone, it would definitely be Claude Opus 4.7.

Gemini

I have a year of Google Pro benefits for Gemini, the student offer Google gave out last year.

Although the Pro benefits last that long, my use of Gemini has dropped a lot over the past two months. Now I mostly use it just for basic questions, to spread out my Claude token usage.

But I have to admit, Gemini's research feature is very good. Give it a research topic and within minutes it produces a detailed, solid report.

And the Google Pro benefits don't apply only to the Gemini LLM. I can also use Google's NotebookLM.

Overall, I don't think Gemini is bad. It's a great researcher and very handy for everyday questions. But for deep conversation and collisions of thought, Claude is still the authority.

ChatGPT

I only subscribed to this today, mainly for the image-2 image generation model updated a couple of days ago.

The past few days, this model has spread widely in AI circles, and its abilities have grown almost mythical in the retelling.

Luckily, you can use the model for image generation in a limited way for free without subscribing. I tried it out, and it really is good. Using the free quota, I redesigned my blog's homepage. Just one draft, and it made my eyes light up.

Let me show the images so you can feel how powerful this model is.

This is my blog homepage's original design:

And this is what the ChatGPT image-2 model designed for me after reading it:

Don't rush to decide which looks better, because taste varies from person to person. What you should really notice is that I used just one sentence as the prompt, "Please redesign the homepage based on my blog," and within a minute I got a usable design.

That's the most striking part. It isn't how capable the model is, or how fast. It's that this kind of capability and change is happening in a form people hardly feel.

Using this method, I redesigned my entire blog from start to finish in no more than ten minutes.

When it was done, I sat in my chair, my mind not even caught up yet, thinking:

That's it?

Yes, the task was finished. But the impact on existing workflows is far from over.

If I hadn't come into contact with AI, or if you set the clock back ten years, how would I get a fully working blog and a design I found acceptable, with no training in art, design, or programming?

I think the answer is that there's no way I could, because I can't even code.

And suppose I could code and had built my own blog. For the design, besides developing my own taste and studying design, I'd have another option: finding a designer.

Finding a designer usually means higher cost and accumulated time.

How long does it take a designer to design a complete blog? I'm not sure. But I know that if I'd hired a human designer, told them what I needed, and they handed me the finished work in under ten minutes, I'd definitely doubt the value of their service.

Yes, in the past time was often how we measured value. But now it's not a person, it's AI, and it can give me a plan in ten minutes. How am I supposed to measure that by time? Because it really did give me a usable design. Its usefulness exists on its own and isn't affected by how long it worked.

So that's the scary part of how AI is developing.

DeepSeek API

I built an agent into my blog. Visitors can chat with it freely within set limits. It's the only piece of work I'm fairly happy with since I started vibe coding.

An agent built into a web page needs a built-in AI API. To make it easy for users in China, I chose DeepSeek's API.

As a pride of Chinese AI, DeepSeek's API is cheap and plentiful, and the quality of its output is very usable. It's been a pleasure to use so far.

MiMo-V2.5-Pro

This is the second model I subscribed to today, a product from Xiaomi.

I'd never used a Chinese model for coding before. Today I saw Khazix's recommendation, so I subscribed to the cheapest Token Plan Lite for thirty yuan.

But before subscribing, I noticed something odd. What the MiMo Token Plan lists isn't a token amount but something called Credits.

Did Xiaomi invent something different from tokens? I didn't think too hard about it. I subscribed first, started using it, then browsed online for a while, and only then realized this model burns through credits fast.

My composure cracked.

In today's real-world test, I found that's exactly the case.

The Lite plan gives me a monthly total of 60,000,000 Credits. By the time I'm writing this, 59% is already used. When it hit 60%, Xiaomi thoughtfully sent me a text message telling me to top up in time.

So what did I use it for?

Really just one thing: adding a feature to my blog's agent Q&A tool that exports chat history as an image, and then designing and revising that image template.

I'd already designed a template with GPT and sent it over, but its reading ability seems poor. It never made a revision I was happy with. The only plus, I'd say, is that it's pretty fast.

It works quickly.

And doing this little bit of work used up 60% of a month's total. I honestly can't afford that price.

So once it's used up, I probably won't keep using it.

That said, Chinese models are now a crowded field too, and there are many I haven't tried, like Kimi, GLM, and MiniMax.

AI is developing like wildfire and changing every day. Nearly every day brings new concepts, new technology, and new anxiety.

At home and abroad, models are everywhere and competition is fierce.

I can't help wondering what the market will look like once AI development settles. Will it be like most industries, where one company with absolute dominance emerges and takes most of the customers?

I can't predict it, but right now I think that's unlikely. Looking at the foreign models, none of them is simply stronger or weaker. Each stands out in its own area of strength. Operating system competition ended up with a few oligopolists (Windows, macOS, Linux), but the restaurant market will always have Sichuan, Cantonese, and Japanese cuisines coexisting, with no "dominant" single choice. AI model competition is closer to the latter, because the reason users choose an AI isn't "which is most widespread" but "which suits my taste best"

After all these years of development, operating systems are down to a few oligopolists (Windows, macOS, Linux). But the restaurant market is always a coexistence of the major cuisines. Nobody dominates the market.

I think AI's development and competition is similar to restaurants.

Because we all choose by taste, not by who has opened the most shops.