Which AI subscription should you choose in 2026?

AI subscriptions are sold as if the model’s power were the only difference. In practice three other things decide, and none of them is on the price list: how the limit is measured, how much you actually pay, and what you get on the cheapest paid plan.

This article is updated in place and collects what we’ve checked on the providers’ pages in the US, with dates. There are no prices here: they change more often than we can rewrite the text, and they’re in the AI category, next to each product.

In short

If you use an assistant every day, the limit matters most — how it’s measured and how much the provider tells you about it in advance.

If you’re planning an exact monthly cost, tax matters: all four providers we checked list prices in dollars, but at least one says outright that its prices don’t include applicable tax.

If you open the assistant once a week, a paid plan probably gives you nothing extra — the free version handles the same thing.

The limit isn’t a number

Of the four assistants we checked in the US — ChatGPT, Claude, Google AI and Perplexity — none promises a fixed number of messages on its price list.

• ChatGPT’s US price list talks about unlimited text chats on Free, but with limits on uploads, image creation, memory and context. On paid plans the same unlimited chats are “subject to abuse guardrails”.
• Anthropic writes under its plans only: “Usage limits apply.” Its FAQ adds that limits reset on a rolling five-hour window and that paid plans add weekly limits on top. The top plan is listed “From” a price — it depends on the multiplier you pick.
• Google says AI Plus gives 2x and AI Pro 4x access to Gemini compared with the free version. How many messages that means, Google doesn’t say.

The conclusion is simple: if the number isn’t on the page before you buy, you won’t get it after you buy either. You learn the limit when you hit it.

The price on the list is rarely the amount on your card

Every one of the four providers lists prices in dollars from the US — but that doesn’t mean the amount on your statement matches.

• ChatGPT’s price list shows dollar amounts per month.
• Google AI’s page shows dollar prices too — and besides Gemini, the plans include cloud storage you can share with up to five other people.
• Claude writes under its plans: “Prices shown don’t include applicable tax.” Depending on your state, sales tax can be added at checkout.
• Perplexity shows its Pro price “when billed annually”; the monthly amount is shown only at checkout.

The cheapest paid plan can have ads

ChatGPT’s cheapest paid plan, Go, “may include ads”, according to the US price list. If you’re paying to avoid ads, read that line first.

What this comparison doesn’t cover

Before you buy, it’s worth finding three things in the terms: what your conversations may be used for, how long before renewal you can cancel, and when you get your money back. These terms differ between providers and change over time, so check the current version on the provider’s own site.

How to choose in five minutes

• What will you do every day — write, work with documents, or look up information with sources? That decides which limit you’ll hit first. Images follow a different logic: Midjourney sells GPU time, in dollars.
• Use the free version for a week. If you haven’t hit a limit, a paid plan gives you nothing yet.
• Check tax: Claude says its prices don’t include it; with the others, look at the checkout total.
• If you’re paying to avoid ads, read the description of the cheapest paid plan.
• Buy where you know how to cancel: a subscription bought in an app is canceled in the app store, with Apple or Google.

What I’ve seen with my own eyes

Everything above comes from what the providers say themselves. This part is about what they leave out: four subscriptions I bought with my own money and used day after day in my own work. The amounts are what I paid myself, in euros — the US prices are on the product pages.

Where to start: the free versions

It all depends on what you want to do with them, and here the line is very clear.

If you just need answers — what something costs, what a word means, how something is spelled — the free version of “ChatGPT” is enough. When its limits start getting in the way, the next step is “ChatGPT Go”, and you don’t need more. For that kind of use I’d skip the pricier plans: you’d be paying for power you never use.

With one caveat: free models sometimes get even simple questions wrong. If the answer is going to be used for something important, check it.

If you want to create — program, write longer texts, work with your own material — the free plan definitely isn’t enough. And it’s about more than the limits.

I’ve tried all three free tiers — “ChatGPT”, “Gemini” and “Perplexity”. None of them shows what the models can really do. Worse: made-up facts show up all the time and are delivered with the same confidence as real ones. In serious work you can’t rely on the free versions. Of the three, I found the free version of “ChatGPT” the most bearable, but the same caveat applies.

The reason is the same mechanism: when the limit is reached, a paid plan switches to a weaker version of the model. On the free plan the weaker version is the permanent state, whether you’ve hit a limit or not.

The paid plans I’ve bought

I’ll start with the cheapest, because that’s exactly the order I chose them in.

“ChatGPT Plus” for €23 turned out to be the best of the cheap plans. The limit is so generous that I’ve never managed to hit it. I can start thorough research, upload a pile of material, work with it all day — and nothing happens. A plan where you don’t have to think about the limit is rarer than you’d think.

“Claude AI Max 5x” (€90, €108.90 with VAT) — this is where a different story begins. The model is really good, but the pool runs out fast, and it’s worth saying how fast. Working in ONE window, without parallel agents and with breaks, I use it up in two days. The five-hour windows are so tight you can’t work through everything in one go: the breaks simply become part of the workday.

And the other end of the same math: with five windows at once, the five-hour limit can run out in fifteen minutes. Same plan, same money — just a different way of working. I’ll come back to that with agents, because that’s where the same thing costs the most.

“ChatGPT Pro” (€103) has much bigger limits, but is less inventive than “Claude”: the solutions come out simpler.

With “Claude AI Max 20x” (€180) I no longer feel the five-hour limits. I’ve still used up the weekly limit in two to three days. The ceiling just moved further away.

“Perplexity Pro” — the opposite experience and the only one of the four I advise against. The answers are short, big questions don’t fit, and the monthly limit for “Computer” mode ran out for me in twenty minutes — yes, twenty minutes. On top of that I never really got along with it: the settings took more patience than I expected. Today’s conclusion is simple: right now I wouldn’t choose it.

About plans I haven’t used myself, I say nothing — they aren’t included here.

The order that pays off

When the previous period ends, I first take “Max 5x”. In a day I use up its entire weekly pool, and the next day I switch to “Max 20x” — the provider prorates the switch by the day and deducts what I didn’t use from the new plan’s price. That way I get the weekly pool of “5x” for the price of one day. If you’re going to end up on “Max 20x” anyway, starting with it is the most expensive route.

This is my own practice. The provider publishes no such rule, and how switching is prorated can change, so check what your own billing window shows before you do the same.

“ChatGPT” almost always agrees

That’s its biggest weakness, and at first you don’t notice it, because it looks like politeness. I ask whether an idea is good — almost always I’m told it is. I suggest a bad solution — I get an explanation of why it’s excellent.

So I take its suggestions with suspicion and check the important ones elsewhere: I show the same answer to “Claude” and simply ask whether the solution is good. The second opinion comes in twenty seconds and has more than once saved me work that would otherwise have had to be redone from scratch.

That’s also why I have both: for the limits, and so one can check the other.

A long window starts losing the thread

Both “ChatGPT” and “Claude” start rambling after long work in the same window. The answers get flat, contradict what we agreed an hour ago, or re-solve what’s already solved. In “Fable” windows it’s especially noticeable.

I don’t wait until the memory is completely full. When I see the window nearing its limit, at about 90 percent, I open a new one without waiting for the last few percent. The only exception: when the work is half done and the window compacts itself — then I keep going, because stopping would cost more.

After every compaction I ask it to read the rules again. When the conversation is shortened, a large part of the agreements disappears — and from the outside it looks like everything is still there. If the work has its own rules file, my first sentence after compaction is always the same: read it again. Otherwise half the rules are forgotten, and you notice only when something is already done wrong.

When you start too many agents at once

This costs money, so it’s a warning.

In tools with agents, it’s tempting to start as many as possible at once. If there are too many and they hit the five-hour limit, it’s not an ordinary pause. When the limit renews, the agents don’t pick up where they left off. They start over from the beginning, as if the earlier hours of work had never happened.

So you pay twice for the same work, and you may get no result at all: credits burned, work abandoned halfway. The practical conclusion is one, and it’s worth sticking to from day one: keep the number of agents strictly in check and split big jobs into parts that fit in one limit window.

I’d skip extra credits

When you hit the limit, there’s always a way out on offer: buy more credits and keep working. I’ve done it a few times, never again.

Credits are disproportionately expensive. For €50 I got so little done it felt like paying for nothing — by my own estimate about twenty times less than the same money would have bought in a subscription. That’s an estimate from my own use; we haven’t measured it. The difference is so big that precision doesn’t matter here.

In practice it gives a simple rule, especially if you don’t have the biggest plan yet: when you hit the limit, upgrade the plan first. Don’t buy credits.

And one thing I say quite seriously, because it’s about money: don’t turn on automatic credit top-ups.

Many tools offer it as a convenience: when credits run out, a charge is made automatically and the work never stops. Put that together with the previous section, and the convenience becomes a trap. Agents that restart after the limit create costs without pause, and the automatic charge removes the last brake — an empty account. Nothing stops it as long as there’s something to charge.

It’s about more than a few dozen euros. In a few days, tens of thousands could disappear that way — that’s my warning from what I’ve seen; we have no documented case. A risk of that size is worth knowing about before you flip that switch.

Three things worth doing today:

• keep automatic credit top-ups turned off;
• if the tool lets you set a hard spending cap, set it right away;
• check the balance yourself after every bigger agent run, and don’t wait for a notification.

What I’d recommend

After all this, two paths remain, and almost nothing in between.

If you just need answers — the free version, then “ChatGPT Go”, and that’s it.

If you create seriously — “Claude AI Max 20x” for €180 and “ChatGPT Pro” for €103. Yes, that’s a lot of money, and I can’t put it more nicely. But if creating is your main work, the cheaper pair stops mid-week, and a stop costs more than the difference between plans. If you create less often, the pair also works with the cheaper “ChatGPT Plus” for €23 instead of Pro — that’s exactly how I work.

And finally a small thing that costs nothing: now and then show the same work to a free assistant too, just as an outside observer. Usually it doesn’t say much — but sometimes it notices exactly what both paid ones missed.

The comparison of the two pairs most often weighed against each other is on separate pages: ChatGPT and Claude and ChatGPT and Gemini.