I pay $200 a month for Claude Max, and I hit 90 to 95 percent of my limit every single week. I built my workflow around it on purpose.
So when a LinkedIn post told me I might be helping to bankrupt an AI company, I paid attention. Christian Guthier wrote that if you use AI every day for everything, congratulations, you might be bankrupting an AI company. Or several at once. Plenty of people shared it, and plenty of them nodded along.
It’s a good line, and the conclusion is wrong. I’m unusual. Most subscribers aren’t anywhere near the limit, and the people telling you Claude loses money on its subscriptions are leaving that part out.
Where does the claim that AI subscriptions lose money come from?
The research firm SemiAnalysis bought every paid plan from OpenAI and Anthropic and ran each one to its usage limit with long, heavy tasks. Then it priced the tokens at the companies’ own pay-as-you-go API rates, the price a developer would pay for the same work.
Those numbers made headlines. Tom’s Guide reported the findings in June: OpenAI’s $200 ChatGPT Pro plan delivered about $14,000 of API-equivalent usage in a month when maxed out, and Claude’s top tier came in around $8,000. SemiAnalysis also estimated where each plan stops making money. For Anthropic’s top tier, it put the line around 10 percent of the allowance, and closer to 20 percent on the cheaper plans.
Read that carefully and you’ll see two problems.
Problem one is the price tag. API rates are retail prices with a profit margin built in. They tell you what a developer pays. The computing costs Anthropic far less. Tom’s Guide said so itself: nobody’s losing $14,000 a month on a subscriber. Price a subscription at retail, call the difference a loss, and you might as well say a restaurant loses money on its own employee meals because they’d cost more off the menu.
Then there’s problem two, and it’s bigger. The test assumed every plan gets used to the limit. Almost none of them do.
Why do AI subscriptions work like gym memberships?
Think about how a gym makes money. People sign up in January, go every day for a month, and then show up once in a while, if at all. The gym keeps billing them. If you count what a membership gives you, it’s underpriced. The price works because almost nobody uses it. I’d guess 95 percent of members don’t. And that’s before you get to how hard the gym makes it to cancel.
Nobody prices a gym by assuming every member lifts weights for five hours a day.
Nobody prices a gym by assuming every member lifts weights for five hours a day. Plenty of products run on the opposite assumption, that customers won’t use what they paid for, or won’t use much of it. AI plans are built the same way.
Leave the free tier out, and look at the people paying. My guess is the average subscriber uses 20 or 30 percent of their allowance. Some weeks they’re busy, some weeks they forget, some months they use it for a recipe and a birthday card. Each of them pays the full price every month.
I’m the member who shows up every day. I’m not typical, and nobody should build a business model around me. The pricing is built around everybody else, and everybody else covers me.
Judging an AI subscription by its heaviest user is like judging a gym by the guy who lives there. – Richard LoweShare on X
So calling these plans money losers is deceptive. Take one extreme customer, price him at retail, and present the result as if it described the whole business. You can make almost any subscription look like a disaster that way.
Does the stranded cruise ship prove flat-rate pricing fails?
Guthier pointed to the MV Odyssey as a warning, and the story is real. Villa Vie Residences sold people a home at sea for a monthly fee, on a voyage that left Belfast in October 2024 and was supposed to run three and a half years. The Associated Press reported that fuel costs per cabin rose from about $1,800 a month to an estimated $6,000 to $8,000 in parts of Asia. The company suspended the voyage in Cambodia, and 203 passengers had to leave the ship, some with seven hours’ notice.
It’s an awful story for the people on board, and it’s a bad comparison for AI.
Every resident on that ship used the whole plan. They lived there. Every cabin burned fuel every day, so when fuel prices tripled, the cost hit every customer at once. There were no light users to carry anybody, and the company had nowhere to hide the bill.
An AI plan works like the gym. Most members barely show up, and the ones who do are a small slice of the bill.
Do AI agents break the gym membership math?
They do, and Grok Bot just showed how.
Grok Bot is the AI agent from Elon Musk’s AI company, launched in beta in August. An agent doesn’t wait for you to type a question. It takes a job, works through it step by step, calls tools, checks its own work and keeps going, sometimes for hours. On October 7, Musk said Grok Bot would hand each task to the best model for the job, naming Claude Opus 5.5 along with Midjourney and Suno.
Then users started hitting their weekly limits almost overnight. On November 1, Musk said the cost of serving Grok Bot was unsustainably high relative to revenue, and some users called the cuts a rug pull.
That’s the exception, and it proves the rule. A chat subscriber is a gym member who shows up a few times a month. An agent is a member who moves in. It never goes home, it never forgets to come, and it runs every machine in the building at once.
No flat price survives a member like that.
Expect agent products to move toward tighter limits and pay-as-you-go billing. The ordinary chat plan most people use is a different animal, and the gym math still holds for it.
How much of your AI plan do you really use?
Most plans show you how much of your allowance you’ve used. Go look.
If you’re at 20 percent, you’re the customer the business is built on. You’re paying for capacity you don’t touch, the same way a gym member pays for the treadmill he walked past in March. Nothing about that is a scandal. Flat pricing works that way, and in exchange you never have to worry about a surprise bill at the end of the month.
Near the top, like me, you’re getting a bargain, and the bargain is paid for by everyone else. I don’t feel guilty about it. I pay full price, I use what I paid for, and I’d guess heavy users are worth more to these companies than the dollars suggest. We push the tools hardest, we find the problems first, and we’re the ones telling other people which tool to buy.
Either way, nobody’s bankrupting anybody.
Could AI subscription prices still go up?
Sure. Any company can raise prices, and AI companies are spending enormous sums on chips and data centers. Energy prices have climbed since the war with Iran choked traffic through the Strait of Hormuz, and power is a big part of what data centers cost to run. I wrote about where all that money comes from in the AI hype cycle.
For subscribers, prices have been moving the other way so far. When Anthropic released Claude Opus 5.5, per-token prices dropped and subscription limits went up. Competition keeps pressure on everybody. If you want to understand what a token is and why everything gets counted in them, my guide to AI terms for writers explains it.
My own line is simple. I’m comfortable at $200 a month. I’m not comfortable with more. If Claude ever charged me the real per-token cost, I’d switch to something else.
Keep your work portable anyway. Your manuscript, notes, story bible and the prompts that work for you belong in plain files on your own computer, outside any one company’s app. Then a price change is an annoyance, not a crisis. I learned the same lesson when Microsoft raised Office prices for an AI I never asked for, and I walked away from Office for a free clone.
My take
Claude isn’t going broke on people like me. I use 90 to 95 percent of my Max plan every week, and the members who use a fraction of theirs cover the difference, the same way they do at every gym in America. Price one heavy user at retail, call it a loss, and you’ve done deceptive math. Use the plan you pay for, keep your files in your own hands, and ignore the doom posts.
Use what you pay for
If you’re paying for an AI plan, the best thing you can do is use it. I built my week around mine, and it’s the best deal in software I’ve found.
For help setting up an AI process that gets real work out of your plan, my AI services page explains how I work, and you’ll find the rest of my AI writing in the AI and writing hub.
The gym is counting on you to stay home. Show up anyway.
For anyone paying for an AI plan and wondering how to get their money’s worth: what changes when you build your work around the tool and keep the thinking for yourself.
