Credit: Wikimedia Commons

We keep hearing that AI companies are burning through money. That sounds pretty strange; they have millions of paying subscribers for services like ChatGPT, Claude, and other AI tools. They use them every day. How can that not be a great business?

Turns out, the math ain’t math-ing.

The problem isn’t the average user asking a chatbot to draft an email. The problem is the heavy user asking an AI agent to write code, search through files, run tools, revise plans, and keep working for hours.

SemiAnalysis, an independent research firm bought every major subscription tier from OpenAI and Anthropic, then pushed the plans to their usage limits with coding and agent-style tasks. The results were striking. A fully used $200 ChatGPT Pro 20x plan could represent about $14,000 of API-equivalent usage in a month. Anthropic’s $200 Claude Max 20x plan reached roughly $8,000 by the same kind of comparison.

Basically, if people would be using their subscriptions to the max, they’d be costing AI companies thousands of dollars.

A Token of Appreciation

Every interaction with a large language model is measured in tokens. A token is a small chunk of text. Sometimes it is a full word. Sometimes it is only part of a word.

The more tokens a model reads and writes, the more computational power it uses. A short question and answer may involve only a few hundred tokens. A long coding session can use vastly more.

This matters starts to matter even more when AI start doing agentic work. Agentic work is when an AI system does more than answer a prompt, autonomously planning steps, using tools, checking results, and iterating toward a goal with limited human guidance. If you have an AI system buy a plane ticket if it goes under a threshold, that’s agentic work; if you have it google stuff and assemble reports then write code on that, it’s agentic work.

×

Thank you! One more thing…

Please check your inbox and confirm your subscription.

Usually, you don’t see much of this work as it happens behind the scenes. But that hidden work is becoming a major cost driver. Even as individual tokens get cheaper, people are asking AI systems to do much more with them. SemiAnalysis reportedly found that some advanced agent-style workflows can consume up to 1,000 times more tokens than a standard chatbot exchange.

For AI providers, that creates an awkward business model. The subscription price is fixed. The cost of serving each user is not.

The Economics Gets Ugly

SemiAnalysis compared flat-rate subscriptions with the cost of buying similar usage through APIs, where developers pay as they go. The gap was huge.

A $20 ChatGPT Plus plan could correspond to roughly $700 in API-priced usage if fully used. The disparity grows at the premium tiers, where the $200 ChatGPT Pro 20x plan could reach about $14,000 in API-equivalent usage. Claude Max 20x, also $200 per month, reportedly reached around $8,000.

The study also examined utilization rates—the percentage of a subscription’s maximum allowance that users consume. OpenAI’s lower-tier plans reportedly become unprofitable once utilization exceeds about 11.4%. For its highest-end offerings, margins disappear at roughly 5.7% utilization. Anthropic appears somewhat better insulated, reaching break-even around 20% on lower tiers and roughly 10% on top-tier plans.

In plain English, a heavy user does not need to fully max out a plan for the economics to become painful.

That suggests heavy subscribers are effectively receiving large subsidies from AI companies. The companies may accept that for now because subscriptions help them attract users, build habits, and compete for market share. But the math gets harder as more people shift from simple chat to complex agent workflows.

Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit. It’s widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing). However, we found… pic.twitter.com/1e0zFhbFuo

— SemiAnalysis (@SemiAnalysis_) June 10, 2026

What Happens Next?

These companies can’t keep losing money forever. But they have few options.

Raising prices is risky. Cutting usage limits is risky too. Users have grown used to predictable monthly fees, and competitors are waiting for any sign of frustration. So what can they do?

The most likely answer is smarter routing. Basically, have different agents of different levels completing different requests.

Simple requests can go to cheaper models. Harder coding, reasoning, or research tasks can be reserved for more expensive frontier systems. Done well, this can cut costs without making the product feel worse for most users.

Others are turning to open-source alternatives such as DeepSeek or building specialized models trained on internal data. In some cases, companies report achieving comparable performance at a fraction of the cost.

There is room for the math to improve. As companies build more data centers and learn to run older AI models more efficiently, today’s advanced tools could become much cheaper to provide. SemiAnalysis suggests that some powerful models may eventually fit into a $20 monthly plan without losing money.

But the newest, most powerful models are a different problem. They may keep costing far more to run than everyday chatbots. If that happens, companies may stop bundling their best features into flat subscriptions and start charging heavy users based on how much AI power they actually use.

For now, the industry faces a balancing act. Users have grown accustomed to predictable monthly fees, while the cost of delivering state-of-the-art AI remains anything but predictable.