My AI setup costs me around 550 euros a month.
That covers ChatGPT Pro and Claude Max, plus hosting and databases. The system behind it does the work of six to seven full-time people and keeps running through the night.
Don't copy that math.
Most of the people who write to me have a different problem. They pay for three or four subscriptions, use two of them halfway, and can't say for sure whether any of them earn their keep. At the end of the month a credit card statement shows up that nobody really reads.
Here's the problem:
Most comparisons stop at ChatGPT, Claude, and Gemini. Coding, image, video, audio, and agents are missing entirely.
This article maps the whole market, gives you four ready-made subscription mixes with a monthly total, and shows you where the limits hit before you have paid for them.
- Four user profiles with a ready-made subscription mix, from $20 to $55 a month
- OpenAI paused new sign-ups for the $200 Pro plan on September 10
- No chat subscription includes video generation anymore now that Sora is gone
- Poe and Perplexity cut their quotas without cutting the price
- Below a few million tokens a month the API is cheaper, above that the subscription wins
1. The short answer: which AI subscription you actually need
Let's start with the answer instead of the table.
If you want exactly one AI subscription, take ChatGPT Plus or Claude Pro. Both cost $20 a month. They cover everyday work, and for images, code, or research you don't need anything else on top at first.
If you work in marketing and push out social content regularly, add a cheap image plan and a voice plan. If you write code, Claude Pro already replaces a separate coding subscription, because Claude Code is part of it. And if you produce video, you won't get around two or three specialist plans, because no chat subscription currently gives you usable video at all.
That was the short version.
The long version is more interesting, because it explains why the prices sit at $20 and at $200 and why almost nothing happens in between. The chat plans show that best.
2. Chat subscriptions compared
On September 10, OpenAI paused new sign-ups for its most expensive ChatGPT plan.
According to a report by Fortune, an employee cited "unprecedented demand" after the launch of GPT-6 Astra. Existing subscriptions keep running unchanged, but nobody new gets onto that tier for now.
That's more than a footnote. The expensive plans look like flat rates, but they work as quotas that a provider can pull back when it runs out of compute.
Here are the chat subscriptions that actually matter.
You probably noticed straight away that there are really only two price points.
As a bar chart the jump gets obvious:
gradually.aiThe entry class sits at $20, and ChatGPT Plus, Claude Pro, and Perplexity Pro are practically tied there. Above that, nothing happens for a long stretch, and then the pro class starts between $100 and $200. A sensible middle tier does not exist, because the providers simply don't sell anything there.
Google prices its plan differently from everyone else. Google AI Pro is listed country by country rather than as one global dollar price, so check what it runs where you are. What you get for it is Deep Research, the Jules coding agent, and limited access to Veo.
Mistral does something nobody else does.
Le Chat Pro costs $14.99 plus tax and hands you a monthly API credit on every tier, including the free one. If you tinker with small automations on the side, that saves you a second account.
Two more plans fall outside the grid, each for a different reason.
SuperGrok from xAI costs $30 and lands between the classes. In return it includes image and video generation, which has become rare at that price.
Microsoft no longer sells its AI as Copilot Pro but as 365 Premium for $19.99. A terabyte of storage and the Office apps come with it. If you pay for Word and Excel anyway, your math here looks different from everyone else's.
One more word about the multipliers in the table.
When a provider advertises 5x or 20x, it always means a multiple of the tier below, not an absolute amount. Neither OpenAI nor Anthropic will tell you how many messages that adds up to in practice. You're buying a factor, not a number, and that is exactly where every comparison table hits its limit.
Three names that show up in every model comparison are deliberately missing here. DeepSeek, Qwen, and Z.ai don't sell a classic consumer subscription. Chat is free at DeepSeek and Qwen, and you only pay for the API or, in Z.ai's case, a pure coding plan. Meta introduced paid tiers in mid-September 2026, but both are still in a limited rollout.
So who gets which recommendation from me?
For text, concepts, and anything that looks like thinking, Claude Pro is currently the strongest $20 you can spend. If you want the widest toolbox in a single subscription, take ChatGPT Plus, where image generation and Codex are already included. How the two differ in daily work is something I go through in my comparison of ChatGPT and Claude. And if you live in Gmail, Docs, and Drive anyway, Google AI Pro is a very good fit, simply because everything sits in one place.
And if you don't want to pay anything right now, that's a decision too. The free tiers have gotten much better than the providers admit, and I go into that in my article on using ChatGPT for free.
Things get more expensive and more confusing with the coding plans.
3. Coding subscriptions compared
This one insight saves you $20 a month.
Claude Code is already part of the Pro subscription for $20. You don't need a separate coding plan, you need the chat plan you may already be paying for. You do share the quota with chat, but for a few hours of work a day it stretches surprisingly far.
I've worked with it since mid-2025 and have spent more than 800 hours in the terminal. Getting started needs no expensive tier.
Two things in this table deserve a closer look.
Cursor and GitHub Copilot bill differently from Anthropic. You're not paying for access, you're buying model usage at list price. The Cursor Pro fee includes $20 of model usage at list price. Copilot Pro+ even drops $70 of credit into your account for a $39 fee.
That sounds like a gift, but mostly it is a different kind of risk. If you work a lot with expensive models, you're done by the 20th of the month. If you barely work at all, you're giving money away.
Anthropic, OpenAI, and Z.ai, on the other hand, name no exact token or prompt quotas per window. I went looking and found nothing. The providers deliberately keep the numbers open and reserve the right to adjust them.
The 20x tier of Claude Max sits at around $200. Anthropic itself only writes "20x more usage than Pro" at that spot on the pricing page, and the concrete figure comes from third-party sources that agree with each other.
Those plans now run under the Devin brand and also start at $20, with the big tier at $200.
With the Chinese providers the data gets thin. The GLM Coding Plan from Z.ai starts somewhere around $10 to $18 a month, depending on which source you read.
Kimi Code is included in all Chinese membership tiers, which are officially listed only in yuan. You won't find a separate international dollar price on the official page. MiniMax stacks its coding plan roughly between $10 and $150 by request volume.
Don't get me wrong:
These plans often cost a fraction of the Western price and are technically solid. You just can't budget properly when the numbers move around, and budgeting is what this article is about.
So my recommendation stays short. Start with Claude Pro, because Claude Code comes with it, and read what that costs in practice in my article on Claude Code pricing. If you'd rather work in an editor than in the terminal, add Cursor Pro, which puts you at $40 together. GitHub Copilot Pro at $10 is the cheapest serious entry point there is, as long as you only want code completion.
Two popular agents are missing from the table, because they have no vendor plan at all. Both OpenClaw and Hermes Agent bill through your own model accounts, so the cost shows up somewhere else.
How that adds up for Hermes Agent is something I break down in a separate article.
For image, video, and audio the math changes completely. There you pay by seconds, images, and minutes.
4. Creative subscriptions: image, video, and audio
4.1 Image subscriptions
Before you book anything here, check your chat subscription.
ChatGPT Plus includes image generation, and for blog images, social graphics, and quick drafts that is plenty for most people. A dedicated image plan only pays off once you need a recognizable style or run through dozens of variations a day.
Midjourney is still the 800-pound gorilla in this round when it comes to image quality. The entry plan costs $10 and gives you a good three hours of fast compute time, and then you're done. If you work longer, you need relax mode, and that starts at the $30 plan. How to get going with it is in my Midjourney guide.
And now something you might not expect from me.
I find Midjourney boring. I sell a course on it, the course is good and the tool is strong, but image generation interests me personally much less than writing and automation. When someone charges you for their own taste, you should get suspicious.
Take Adobe Firefly when your images have to be legally clean for client projects. Leonardo and Krea cost less, and both bill in token or compute quotas. Ideogram Basic currently runs around $7 to $10 a month, depending on which promotion is live.
One name has changed. Freepik has been called Magnific since late April 2026, with the same company and the same billing behind it, and the old address now redirects.
Magnific sells its credit as an annual package, which works out to roughly 130 images a month including a commercial license.
4.2 Video subscriptions
Here is the news that a lot of comparisons still haven't caught up with.
Sora is gone. OpenAI announced the shutdown in March 2026, the app and the web interface went offline in April, and the API only runs until the end of September. That means no chat subscription includes full video generation anymore.
This changes the math for everyone who needs video. You can no longer cover that need on the side with your chat plan, so you either book a second subscription or you skip it. My test of the current AI video generators gives you an overview of the best tools for the job.
With video, a number other than the price decides your plan, namely the seconds.
Runway Standard costs $15 and delivers roughly 52 seconds of video a month for it. Less than a minute. The Pro tier at $35 gets you a good three minutes. If you publish video regularly you will slide into a higher tier fast, and that is something to know beforehand rather than notice on the fifth day of the month.
Google AI Pro brings Veo along in a stripped-down variant, but with a daily cap. For a few clips a month that works, for production it doesn't.
Kling sits at roughly $7 to $10 for the entry tier according to several sources that disagree with each other. What the free plan is not allowed to do matters more than the exact amount anyway, and that comes up in the chapter on hidden costs.
4.3 Audio and music
The cheapest sensible spend in this whole comparison is $6.
That's what the entry tier of ElevenLabs costs, and only from there are you allowed to use the generated voices commercially. For roughly 30 minutes of audio a month that's enough, for a weekly podcast it isn't. Then you land on the Creator tier at $22. All the details on the tiers are in my ElevenLabs pricing breakdown.
Suno Pro costs $8 and allows 20 song downloads a month. For intros, background music, and jingles that is solid, and it is one of the few subscriptions in this comparison where I consider the quota generous enough.
The next category is the youngest. And the messiest.
5. Agent and productivity subscriptions
This is the first point where things get uncomfortable.
The agent category is the weakest of all four. Not because the tools are bad, but because almost all of them bill in credits that nobody can convert into working hours. You buy 4,000 credits and have no idea whether that covers two tasks or two hundred.
Manus and Genspark are the best-known autonomous agents with a consumer plan, and both sit around $20 to $25. Lovable and Bolt target people who want to build small apps without writing code themselves. Otter and Granola take meeting notes, and at $14 per user Granola is the cheapest way to automate that, with more options in my overview of AI meeting assistants.
Notion AI is a special case. Since May 2025 the AI features are no longer available as a standalone add-on, only inside the Business plan. So if you use Notion privately and want the AI on top, you pay for a jump to a whole new plan tier.
And then there's Poe.
My advice for this category is blunt. For the vast majority of readers, not a single subscription here is worth it. Book one only when you have a concrete, recurring task for it to take over, for example meeting minutes or prototypes. Everything else you can handle with your chat subscription, and without counting credits.
Now for the question most comparisons never get to.
6. Which subscription mix fits you
Nobody subscribes to exactly one tool from one category.
That is precisely where most comparisons stop. They list prices per provider and leave you alone with the question of what the combination costs at the end of the month. So here are four typical cases, worked through with real list prices and no wishful thinking.
The totals are lower than you probably expected.
As a bar chart it looks like this:
gradually.aiThe casual user stays at $20 and needs nothing else, because ChatGPT Plus already brings image generation along. The solo marketing freelancer lands at $36, because a cheap image plan and a voice plan come on top.
The developer pays $40 for Claude Pro and Cursor Pro, so for two different ways of working on the same problem. Only the video producer reaches $55, because nobody sells them video inside a chat subscription anymore.
Not one of these profiles needs a $100 tier.
The pro plans only pay off once you spend several hours a day in a tool five days a week and keep running into limits. You will notice that on your own, and very clearly.
One question comes up constantly anyway, as soon as things get tight.
Second subscription or higher tier?
The answer depends on what you're failing at. If you hit limits although the tool does everything you need, you go up a tier. If a capability is missing instead, meaning video, voice, or a second model provider, you book the second subscription. Don't mix those two up. Otherwise you pay $100 for a problem that $15 would have solved.
For context, here is a number from the outside. In October 2025, the payment provider Bango surveyed roughly 2,000 Americans with at least one paid AI subscription for its report "The Rise of the AI Subscriber".
According to that survey, the average AI subscriber pays around $66 a month. Almost a quarter spend more than $100. Bango sells subscription bundling itself and therefore has a commercial interest in high numbers, which belongs in the picture.
The number still works as a warning sign. All four profiles in my table come in below it, some of them by a wide margin.
Before you click buy anywhere, let's talk about the costs that appear on no pricing page.
7. The hidden costs that no pricing page shows
The price is only half the bill. The other half sits in the fine print, and almost nobody reads that.
7.1 "Unlimited" does not mean unlimited
In February 2026, Perplexity slashed the quota for Deep Research. Instead of roughly 500 requests a day, about 20 a month were left, with no price change and no announcement.
A paying annual subscriber documented publicly that the pricing page kept advertising unlimited research while his quota had long been cut. In the end he filed a chargeback through his credit card.
Poe did something similar, only more quietly.
And yes:
Both providers are allowed to do this, because nobody can hold a marketing phrase to account when it is not a contractual commitment. That is exactly why this chapter exists.
7.2 List prices are pre-tax
Almost every price in this article is a pre-tax list price. Anthropic states on its pricing page that applicable taxes are not included, and Mistral says the same.
So the number on your statement is rarely the number on the pricing page.
7.3 Credits that simply expire
At Genspark, unused monthly credits expire at the end of the billing cycle. With the coding tool Kilo Code this hits the monthly bonus credits, while separately purchased credit stays put according to the provider. Plenty of people miss that difference, and it decides whether paying a year upfront is worth it.
My rule of thumb here is simple. Never pay for credits upfront to cover a need you are only imagining.
7.4 Commercial rights are not in every plan
This is where the most expensive mistakes hide.
The free Kling plan caps you at 720p, stamps a visible watermark into the frame, and explicitly allows private use only. Only the first paid tier gives you watermark-free exports and commercial rights. At ElevenLabs the commercial license likewise starts with the cheapest paid plan.
Runway goes a step further. Every output carries a C2PA provenance watermark regardless of plan, and outside of Enterprise, Runway trains its models on what you upload by default.
7.5 Annual billing saves money and locks you in
Nearly every provider in this article gives you a discount for paying twelve months upfront. Claude Pro drops from $20 to $17. Runway Standard slides from $15 to $12, and at Otter the price roughly halves.
Sounds like simple math.
However:
What you're really doing is placing a twelve-month bet in a market where providers cut quotas and switch products off. Play.ht was discontinued entirely at the end of 2025, and the accounts and cloned voices were deleted with no migration path. Sora and Windsurf also disappeared inside a year.
So I keep it tight.
Annual billing is only an option for the one subscription you have used daily for at least three months. Everything else stays monthly, even if that leaves a few dollars on the table.
And then there is the route with no subscription at all.
8. When the API beats a subscription
I get this question constantly, usually from people who have just seen their first monthly bill.
The idea behind it sounds reasonable. Instead of a flat fee you pay only for what you actually consume, per token. So let's work it through.
I'm putting one usage pattern on the table that is typical but stays an assumption, namely 80% input and 20% output tokens. For Sonnet 5 that gives a blended price of $3.60 per million tokens.
That puts the break-even for Claude Pro at roughly 5.6 million tokens a month. Use less and the API is cheaper. Use more and the subscription wins. My API cost calculator lets you run the same math for your own consumption.
For the 5x Max tier the line moves to just under 28 million tokens. The 20x tier starts paying off at around 56 million. For GitHub Copilot Pro+ the break-even sits at roughly 10.8 million.
Sounds like big numbers? They are.
Anyone working several hours a day with a coding agent reaches those volumes without much effort. For you that means the subscription almost always stays the cheaper option, and the math is settled.
For everyone else the reverse applies. If you ask a question twice a week, the API might cost you $2 a month instead of $20.
The API does come with a catch that shows up in no break-even calculation. You get no app, no projects, no web interface, and no voice mode, you get an interface. So you need a tool that sits on top of it, and that opens a whole new construction site.
Here is what it looks like on my side.
9. What I pay myself
I have two AI subscriptions, and both sit in the most expensive class.
ChatGPT Pro and Claude Max. Hosting and databases come on top, and together that runs to around 550 euros a month. For that I get a system that takes the work of six to seven full-time people off my hands, drafts newsletters and articles in my voice, sorts my inbox, helps keep the books, and watches rankings.
Before you set up the same math, a few caveats.
First, hosting and databases are part of those 550 euros, so these are not pure subscription costs. Second, this system grew over many months, not over a weekend. The first 40% to 50% of the automation took me the longest by far.
And third, my role has shifted. I barely write myself anymore, I review, decide, and sign off. That role justifies the expensive tiers. A normal working day with three ChatGPT questions does not.
What I don't pay for is at least as telling.
I have no video subscription, because I don't produce video. I have no agent subscription, because my automations run through the models I already pay for. And as I said, I find Midjourney boring, my own course on it notwithstanding.
That is the real point behind this whole article. My combination fits my work and would be waste for almost anyone else.
Which leaves the question of what you should do now.
10. Verdict: how to decide
If you do exactly one thing after this article, do this.
Book a single $20 subscription and use it consistently for 30 days for everything that comes up. Take ChatGPT Plus if you want a broad toolbox. Claude Pro fits if you write or code a lot. And take Google AI Pro if your work already lives inside Google's apps.
After 30 days you know two things that no table can tell you. You know whether you regularly hit limits at all. And you know which task your subscription does not handle.
That is what you book the second subscription for. Not earlier.
This sounds slower than it is. In my experience it is the only route that doesn't end in a stack of half-used subscriptions you haven't looked at in months.
And yes:
Cancelling is completely fine. In the Bango survey, more than half of subscribers said they end AI subscriptions as needed and start them again later. That is not failure. Cancelling is a sensible reaction to a market that reshuffles itself every few months.
Which model actually fits your task is a different question from the price, by the way, and I go deeper on that in my comparison of AI models.
And if the whole subscription question felt too granular, my list of the best AI tools by use case gives you a different angle on the same market.
