A $20 Monthly AI Subscription: Are AI Companies Actually Paying Hundreds of Dollars a Month for You?

Imagine this: After work, you visit a newly opened, premium all-you-can-eat Wagyu beef restaurant in your neighborhood. The entry fee to enjoy unlimited premium beef is a mere $20 (about 27,000 KRW). Excitedly, you keep ordering more meat. But in reality, the actual cost of the meat you consume during a single meal is well over $100. The restaurant owner cries tears of blood inside every time you ask for another plate, but they continue to serve the premium meat while taking massive losses, all to attract customers and establish themselves as the best dining spot in town.

Did you know that exactly the same thing is happening behind the scenes of the massive artificial intelligence (AI) services currently shaking up the global tech industry? Behind the magical AI assistants we use for $20 a month—or $100 for professional use—hides an absurd “loss-making competition” among corporations. According to an in-depth analysis by an IT strategy and architecture publication, for every $100 heavy users pay, AI companies might be pouring in more than $1,000 to keep their servers running Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them – R&A IT Strategy & Architecture.

How long can this precarious, unrealistic party continue? And on the day the “owner” finally declares, “From now on, I will charge you for every single slice of meat you eat,” how will our daily lives change?

Why It Matters

We are highly accustomed to binge-watching dramas all weekend on Netflix or leaving music playing all day on Spotify. The flat-rate subscription model, where you pay a fixed monthly amount for unlimited access to content, is considered absolute common sense in modern digital society. This is because once video or music files are uploaded to a server, the additional cost incurred by the company approaches zero, even if millions of people stream them simultaneously. To use an analogy, it’s like heating up a pre-cooked meal in the microwave and serving it to multiple people. Therefore, many people vaguely assume that AI services like ChatGPT or Claude operate on the same structural basis.

Put simply, artificial intelligence is on a completely different level from playing a pre-recorded video. AI doesn’t send you a fixed file; every time you ask a question, massive computers in far-off data centers must consume enormous amounts of electricity to solve math problems in real-time and “think” up a response. The longer the question and the more complex the required answer, the more electricity and computational resources the AI consumes, growing exponentially. Metaphorically speaking, it’s as if a personal chef receives a fresh ingredient in the form of your new question and cooks a completely unique dish from scratch every single time.

The Explainer

To properly understand this bizarre deficit structure, you need to know how AI reads and writes our language and incurs costs. Two key concepts emerge in this process.

The first is the ‘Transformer’ (an AI architecture that understands the relationships between words in a sentence). This isn’t a robot toy that turns into a car or a plane, but an AI brain structure model that grasps context from vast amounts of text data and predicts the next word. The second is the ‘Token’ (the basic unit AI uses to read and write text). AI does not recognize our sentences as a whole like humans do; instead, it digests words by breaking them down into tiny puzzle pieces called “tokens.”

Let’s use an analogy. Imagine you are traveling in a taxi. The taxi meter ticks up the fare every time the wheels turn and cover a distance. In the AI world, the meter ticks once when the AI reads the puzzle pieces of your question (input tokens) and ticks again every time it writes down new puzzle pieces to generate an answer (output tokens). This entire process of asking the AI a question and receiving an answer is called “Inference.”

If you look at the pricing for APIs (intermediaries that help operating systems or programs communicate) used by businesses rather than general consumers, you can feel just how frighteningly fast this meter runs in reality. In the case of Anthropic’s highest-performing model released in May 2026, ‘Claude Opus 4.8’, it possesses the overwhelming ability to remember a staggering 1 million tokens at once—equivalent to dozens of thick books. However, it also comes with massive operational costs: $5 per 1 million input tokens and $25 per 1 million output tokens Anthropic API Pricing in 2026: Complete Guide — Models, Caching, Batch & Optimization.

To make matters worse, if security or regulatory requirements—such as those for financial data—dictate that you must route through US-exclusive regional servers on cloud platforms like Amazon Bedrock or Vertex AI, an additional 10% premium surcharge is tacked on. For example, a team using 100 million input tokens and 30 million output tokens a month would see their base fee of $1,250 jump to approximately $1,375 after applying the regional surcharge Anthropic API Pricing: Claude Opus 4.8 Costs Explained - Amnic.

Costs literally explode when users push the AI “harshly,” such as when software developers write code. Throwing thousands of lines of code entirely at an AI during the programming process and brute-forcing operations to fix bugs consumes an unimaginably massive amount of tokens. Some analyses point out that to support such complex coding tasks for power users paying $100, the company might be eating more than $1,000 in server costs behind their backs. They raise a fundamental question: “Even if we paid the true cost, would it still be profitable to use these tools to reduce the number of programmers?” Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them – R&A IT Strategy & Architecture.

Prominent developer Simon Willison also shares the mood on the ground. He noted, “Stories are frequently circulating about companies that let employees use AI freely on the company dime, only to be shocked by unimaginably massive LLM billing invoices,” adding that this proves OpenAI and Anthropic have completely infiltrated the enterprise market I think Anthropic and OpenAI have found product-market fit. Furthermore, when you consider the massive electricity bills incurred by running graphics cards at full capacity to power lightweight open-source models on a personal computer, it is hard to even guess the level of power consumption by these massive centralized corporations I think Anthropic and OpenAI have found product-market fit ….

Where We Stand

Even in this anomalous cost structure where the company’s cash burns away as servers run, companies are absorbing the immediate losses. This is because the competition among giants to seize hegemony in the AI market is heating up more fiercely than at any other point in history.

Frontrunner OpenAI, founded in 2015, secured massive investment in early 2026, gaining an astronomical valuation of approximately $852 billion. They dominate the global consumer AI market with an overwhelming 900 million weekly active users via ChatGPT Anthropic vs OpenAI 2026: 30x Revenue Gap and 4x Context …. Massive capital, equivalent to a national budget, is solidly backing their ongoing deficits.

Not to be outdone, rival Anthropic’s momentum is equally fierce. Founded in 2021 by former OpenAI researchers, Anthropic has been developing technology based on the firm philosophy of building AI systems that are “steerable, interpretable, and robust without fatal errors” Anthropicisthe new AI research outfit fromOpenAI’s Dario Amodei…. They emerged as a powerful rival to OpenAI by successfully launching the Claude 4 and 4.1 models, equipped with robust reasoning and highly advanced coding capabilities Anthropic API Pricing in 2026: Complete Guide — Models, Caching, Batch & Optimization.

However, behind this glamorous spotlight and flood of investment, a desperate struggle for survival continues. To offset these massive server costs even slightly, OpenAI has rolled out a $100/month Pro plan, and Anthropic has similarly launched a $100/month Max plan, targeting professional groups with the ability to pay I think Anthropic and OpenAI have found product-market fit. The unfortunate dilemma, however, is that power users who purchase these expensive plans squeeze the AI models to their limits to generate code and perform reasoning, meaning the actual computing costs the company must bear increase much steeper than the price hikes. They are revenue-generating customers, but paradoxically, they are also the entities burning through the company’s servers the fastest and accelerating deficits.

What’s Next

What all IT analysts and economic experts uniformly warn about is clear: “This sweet all-you-can-eat party can never last forever.” They cannot keep pouring water into a bottomless vessel.

The current subscription business model cannot be permanently sustained when users push models to their limits, and companies nearing their profitability ceilings will gradually have no choice but to undergo a major overhaul of their pricing systems. As developers and giant corporate clients are already experiencing, preparations are underway to fully introduce a “Usage-Based Billing” system for general subscribers as well, where they pay exactly for the “tokens” they use, much like a taxi meter Usage-Based Billing, No Flat Rate: Why Anthropic’s 2026 ….

However, the ultimately inevitable reality is approaching. In the near future, we will awaken from the illusion of unlimited pricing plans, and a cold re-evaluation will take place regarding whether the value provided by artificial intelligence truly exceeds the “actual cost” we must pay.

AI’s Take

The pleasant, cheap unlimited pricing plans we enjoy right now are not simple acts of charity. They can be seen as a highly calculated “strategic deficit” intended to make the public’s daily lives entirely dependent on AI. Until the habit of asking AI questions and delegating complex tasks anywhere, anytime becomes deeply rooted in modern life, tech giants are willingly absorbing massive bills on our behalf.

But the day will inevitably come when these sweet subsidies are cut off, and the actual monetary amounts corresponding to our true usage will be stamped onto our bills. When that time comes, we will seriously reconsider the weight of a single casual request we casually throw out every day, such as “Rewrite this to be more fun.” Are we truly ready to pay out of pocket for the convenience of the answers AI provides? The true test we are about to face is not a question of how smart the technology gets, but whether AI is essential enough to our lives to make us open our wallets.

References

1. Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them – R&A IT Strategy & Architecture

2. [Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them

   Hacker News](https://news.ycombinator.com/item?id=48434342)

3. [AI’s Spending Problem: Why Fixes Hurt OpenAI, Anthropic

   StartupHub.ai](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/ai-s-spending-problem-why-fixes-hurt-openai-anthropic)

4. Anthropic API Pricing: Claude Opus 4.8 Costs Explained - Amnic
5. Anthropic API Pricing in 2026: Complete Guide — Models, Caching, Batch & Optimization
6. I think Anthropic and OpenAI have found product-market fit

7. [OpenAI and Anthropic Launch Price War for AI Coding Tools

   KuCoin](https://www.kucoin.com/news/flash/openai-and-anthropic-launch-pricing-war-for-ai-coding-tools)

8. SF Victorian AcceptsAnthropicorOpenAIStock asPayment

9. [A San Francisco seller wantsOpenAIorAnthropicstock for…

   Fortune](https://fortune.com/2026/06/03/san-francisco-homeowner-wants-anthropic-openai-stock-for-3-million-dollar-home/)

10. Anthropicisthe new AI research outfit fromOpenAI’s Dario Amodei…
11. Anthropic vs OpenAI 2026: 30x Revenue Gap and 4x Context …
12. Usage-Based Billing, No Flat Rate: Why Anthropic’s 2026 …
13. I think Anthropic and OpenAI have found product-market fit …
14. Anthropic files for IPO before OpenAI as trillion-dollar …