I Cancelled 4 AI Subscriptions Last Month — Here's the One I Kept

Last Wednesday night, I was doomscrolling in bed when a message popped up on WeChat (China's everything-app).

"I can't take it anymore," my friend wrote.

His browser tabs looked like a UN general assembly — ChatGPT drafting emails, Claude rewriting copy, Gemini analyzing images, Grok monitoring trending topics. Four tabs, four subscriptions, and every monthly charge made him wince.

Then came the question: "Which one should I keep?"

My response: "First, tell me what you actually do."

Cross-border ecommerce. Product descriptions, customer emails, competitor analysis, occasional marketing copy.

"You don't need four," I shot back.

Seriously. He doesn't.

This whole thing reminded me of myself last October. I had five — count 'em, five — subscriptions running simultaneously. GPT. Claude. Gemini. Grok. Plus a Chinese model. Opened my phone and felt like an AI arms dealer. So professional. So cutting-edge.

Then the monthly bill hit.

Nearly gave me a heart attack.

It was approaching my rent. And I don't live cheap.

Total disaster.

Spent two weeks running every model through my actual workflow. Kept what worked. Slashed the rest. I've been using the same setup for almost a year now, and honestly? It's comfortable.

At this point you're probably wondering: so which one survived the purge?

Hold on. Let me walk you through each one.

ChatGPT: The Reliable Workhorse (With One Annoying Habit)

ChatGPT is my daily driver. Not because it's the best at everything — though it's pretty damn good — but because it's the most stable.

Stable to what degree? Daily writing, translations, research compilation — it almost never makes me want to throw my keyboard across the room. Think about it: a tool you've used for a year that hasn't made you snap. That's rare.

But there's this one thing. Still unfixed.

It doesn't sound human.

Really. Ask it to write a product description, and you get: "This product leverages cutting-edge craftsmanship to deliver an unparalleled experience." I'm like, can you speak like a person? It says yes, then gives me: "This product utilizes advanced technology to present exceptional sensations."

...sure. Thanks.

So here's my current workflow: ChatGPT drafts, Claude polishes. Claude's writing — I should say Claude's Chinese writing — flows far more naturally than GPT's. At least it reads like a human wrote it, not some "Dear Valued Customer" plastic fantastic nonsense.

But Claude has problems too.

Big ones.

Claude: The Brilliant Liar

Back in the Opus 4 days, Claude's coding ability was genuinely untouchable. In my dev circles, we used to say: if you're coding without Claude, you're making life harder for yourself. Period.

Then came the later versions. Started "drooling."

That's our insider slang for when a model starts spouting complete nonsense. You ask about X, it answers about Y — with absolute confidence, like it's reciting gospel truth.

And it got weirder.

The Chinese language ability improved dramatically, but information retrieval? Abysmal. Ask it a simple factual question and it'll fabricate a whole story — dates, locations, people, statistics — all laid out with crystal-clear conviction.

All of it fake.

Last month I asked Claude to check a competitor's price range. It gave me three numbers. I nearly dropped them straight into my report. Something felt off — thank god for that instinct — so I verified manually.

Every. Single. Number. Wrong.

Unbelievable.

And Claude is expensive. Pro version runs $20/month, and free quota? Gone in two minutes. I've got a friend using Sonnet without paying — he says every interaction feels like racing against a countdown clock. One wrong prompt and you're locked out.

Let that sink in: $20 a month for an assistant that confidently lies to your face.

What do you even do with that?

Gemini: From Hero to Zero

Honestly? My feelings about Gemini are... complicated.

The Pro era? Excellent capabilities, solid world knowledge, strong multimodal features. Basically unlimited image uploads. Daily chatting felt smooth. I used it constantly back then, genuinely thinking Google had finally gotten their act together.

Then Google started slashing.

They nerfed Flash into a dimwit. Flash-Lite became a half-wit. The latest Flash version — supposedly updated recently — has devolved into... look, I'm not the only one using that word. Multiple people said the same thing.

Can't handle basic arithmetic.

Think about that. An AI that can do calculus but can't compute 37+58.

Let that marinate for a second. It's almost surreal.

There's a bright side though: the free tier is generous. Flash-Lite is wide open, images mostly unrestricted. So I use it for exactly one thing now — image generation. Its drawing capability is surprisingly good, better than most alternatives, and there aren't heavy copyright restrictions. Generate whatever you want.

Social media and entertainment? Gemini's your buddy. Actual work? Stay away.

Remember that.

Grok: My Most Regrettable $10

Grok.

sigh.

I subscribed for one month. Ten bucks.

That $10 represents my worst subscription decision this year.

Aside from slightly looser content restrictions — and I mean slightly, not anything-goes — it's nearly useless. Information retrieval is weaker than GPT. Writing feels painfully stiff. Model hallucination is intense. You ask it to understand a simple request, and it interprets it into something from a parallel universe.

Elon hyped the hell out of it — real-time web access, exclusive X data. In practice? Whatever it surfaces, I can find on Google too. Usually more accurately.

Ten. Dollars.

You know what else costs $10? Two fancy bubble teas. And those don't gaslight you.

Definitely cancelling before next month's charge.

DeepSeek: The Dark Horse

This one surprised me the most.

Last year's V3 version? Honestly, not great. Hallucination city. Endless rambling. Obvious capability gaps. My review at the time: no advantage except being free.

But the V4 release this year —

Completely different animal.

Current DeepSeek is essentially a compact GPT. Its Chinese language logic and writing quality? Best among all models — foreign ones included. I've done side-by-side comparisons, and for Chinese tasks, DeepSeek's language intuition feels the most natural. It genuinely understands how Chinese speakers communicate.

Some expressions Claude can't nail. GPT can't nail. Only DeepSeek gets them right.

And it's cheap. API pricing runs about one-tenth of GPT — $10 worth of credits lasts forever. I connected the API through Cherry Studio (a desktop client), and both performance and rate limits beat the web version noticeably.

Very comfortable.

Obvious downside: still no multimodal capability. Can't generate images. Can't recognize images. If you need those features, you'll need other tools.

But let's be real — at that price point, what are you expecting?

So Which One Should You Keep?

If you're like my ecommerce friend and want just one primary model, I'd recommend ChatGPT. It's not necessarily #1 at everything, but it's the most well-rounded and makes the fewest mistakes.

Budget-conscious or mainly doing Chinese content? DeepSeek V4 is more than sufficient — and in some areas, it's actually better.

Coding? Claude Opus 4 — I'm talking the older version, not the newer ones — remains many programmers' go-to. But you'll need to accept the price tag and occasional "drooling" episodes.

Deep in Google's ecosystem? Keep Gemini's free tier. Lightweight tasks and image generation only. Don't give it anything important.

As for Grok...

Let's just not.

The Real Lesson Here

Here's something I realized while writing this.

Everything I just said might be obsolete next month.

This space moves ridiculously fast. A model that was solid last week might get nerfed tomorrow. Last year's undisputed leader could become thoroughly average. I'm writing this in June 2026 — by the time you're reading, version numbers have probably shifted again.

So instead of telling you "use this one," here's what I actually want to say:

Don't get attached to any single model.

They're tools — not religions. Today's favorite might betray you tomorrow. Today's underdog might dominate next version.

The ability to switch — that's more valuable than picking the "right" model.

What do you think?

Tools get outdated. The ability to choose doesn't.

TL;DR / Key Takeaways

• ChatGPT: Best all-rounder. Reliable but sounds robotic. Use for drafts.
• Claude: Great coding (Opus 4), great writing, but expensive and confidently lies about facts. Verify everything.
• Gemini: Free tier is generous. Use for image generation only. Don't trust it with logic.
• Grok: $10/month. Just... don't.
• DeepSeek V4: Surprising dark horse. Amazing Chinese output. Super cheap API. No multimodal yet.
• The real skill: Don't marry a model. Stay flexible. Switch when needed.

What's your AI subscription stack looking like? Found any hidden gems or overhyped disappointments? Drop a comment — I genuinely want to know what's working for you.

AI #Productivity #TechReview #ChatGPT #DeepSeek #Gemini #Claude #Grok