Charging $200 a month but potentially losing $14,000, the money-losing business behind ChatGPT and Claude has been exposed

"$200 per month, unlimited usage."

In the past two years, this has been one of the most successful business models in the AI industry. From ChatGPT to Claude, large model providers have quickly attracted a large number of global users with relatively fixed and predictable subscription prices, packaging the originally complex API billing model into a consumer product acceptable to the general public.

However, a recent analysis report from research institution SemiAnalysis has revealed the less glamorous side behind this model: If users fully utilize the capabilities of high - end subscription packages, OpenAI and Anthropic are likely to face severe losses.

Calculated according to API prices, a ChatGPT Pro user who pays a monthly subscription fee of $200 could theoretically consume model resources worth up to $14,000; for Anthropic's Claude Max package at the same price, the corresponding theoretical cost is also close to $8,000.

A monthly fee of $200 could be worth up to $14,000 when fully utilized

The data that has attracted attention this time comes from a test by SemiAnalysis on the subscription systems of OpenAI and Anthropic.

The researchers conducted extreme stress tests on multiple paid packages of both parties. They simulated the real - world usage scenarios of heavy users, including long - term programming tasks, complex Agent workflows, multi - round reasoning tasks and other high - load applications, and continuously invoked the models until they reached the usage limits set by the platform.

Subsequently, the researchers converted these invocation volumes according to the official API pricing standards. The results were surprising:

● The monthly fee for OpenAI's ChatGPT Pro 20x package is $200; if users fully utilize the quota, calculated according to API prices, the corresponding cost could reach up to about $14,000.

● The situation with Anthropic is similar. For the Claude Max 20x package priced at $200, under the theoretical maximum usage, the corresponding Token cost is about $8,000.

So in extreme cases, the subscription fees paid by users are even less than 2% of the platform's actual costs.

Of course, in reality, most users will not exhaust their quotas, so these numbers mainly represent the theoretical upper limit. But it also shows that for AI companies, what really determines profitability is not the subscription price, but the actual utilization rate of users.

OpenAI faces greater pressure than Anthropic

For this reason, SemiAnalysis further calculated the break - even points of different packages.

The results show that Anthropic's current business model is relatively healthier: For its Claude Pro and Claude Max 5x packages, when the user utilization rate reaches about 20%, it can basically achieve break - even.

In contrast, OpenAI's situation is not so optimistic: When the utilization rate of ChatGPT Plus and ChatGPT Pro 5x users exceeds 11.4%, OpenAI will start to incur losses, and the situation for higher - end packages is even more serious.

The researchers pointed out that the gross profit margin of Anthropic's top - tier package will drop to zero at a utilization rate of about 10%; for OpenAI, a utilization rate of only about 5.7% will put it into the negative gross profit range. That is to say, for some heavy users, they don't even need to continuously invoke the model for 24 hours. As long as they maintain a relatively high usage frequency in the long term, the platform will basically be in a state of "the more services provided, the more losses incurred".

OpenAI CEO Sam Altman has also publicly admitted that the company is facing this contradiction. As Token consumption continues to grow, cost pressure has become a serious problem. OpenAI is currently working hard to find ways to enable users to obtain more value at a lower cost.

The root of all these problems lies in the changing usage patterns of AI.

In the early days, the typical scenario for ChatGPT was simple Q&A, where users would ask a question and the model would return an answer. This mode consumed a relatively limited number of Tokens. But now, more and more users are starting to use AI programming assistants, multi - step Agents, long - context analysis systems, automated workflows, etc. These tasks generally require the model to think repeatedly, plan, invoke tools, and generate a large number of intermediate results.

SemiAnalysis pointed out that in some Agent scenarios, the Token consumption could even reach 1000 times that of ordinary Prompts.

Microsoft, Meta, and Amazon have started to control internal AI expenses

As employees use AI tools extensively, corporate bills have begun to swell rapidly. Previously, Microsoft, Meta, and Amazon promoted the widespread use of AI among employees, but after the costs continued to rise, they have started to tighten relevant policies.

One widely circulated case is particularly noteworthy: A company burned through $500 million on Anthropic's services in just one month because it did not restrict employees' access to Claude.

Facing high costs, many companies are adopting a more pragmatic strategy: dynamically switching models according to task difficulty. The Wall Street Journal previously reported that in this way, companies can save up to 95% of AI costs.

To put it simply:

● Complex reasoning tasks are assigned to cutting - edge models such as GPT - 5 and Claude Opus;

● Routine summarization, classification, information extraction and other tasks are handled by lower - cost models.

As Vishal Misra, the deputy dean of the School of Engineering at Columbia University, said: "Not every task requires a model that understands quantum gravity." In his view, the capabilities of open - source models are rapidly improving, and the possibility for AI providers to charge high premiums based on top - tier models will become smaller and smaller.

It is understood that some startups have taken the lead in making this transformation.

Flo Crivello, the founder and CEO of AI assistant company Lindy, recently said that the company has migrated 100% of its traffic to DeepSeek V4 and no longer uses Anthropic models. The reason is simple: After a large number of tests, the team believes that DeepSeek V4 can achieve results close to those of Claude Sonnet in most scenarios, and the cost is much lower. Crivello said that this migration has saved the company millions of dollars.

In fact, not only DeepSeek, but also open - source model camps such as Llama, Qwen, and Mistral are growing rapidly. When the gap in model capabilities continues to narrow, cost advantages often become the decisive factor.

Will AI costs decrease in the future?

The good news is that the industry generally believes that in the long run, AI inference costs will continue to decline.

With the increase in GPU supply, the expansion of data centers, the optimization of model architectures, and the improvement of hardware efficiency, the operating costs of mid - range models are expected to further decrease. SemiAnalysis predicts that in the future, a model with the capabilities of Claude Opus 4.8 could theoretically achieve profitable operation at a monthly cost of about $20.

However, this does not mean that all models will become cheaper. For the most advanced cutting - edge models, the situation may be the opposite. These models have the strongest reasoning ability, the longest context, and the most complex Agent capabilities, and their training and inference costs are still very high.

Therefore, more and more analysts believe that in the future, the capabilities of top - tier models may no longer be packaged into unified subscription packages, but will be sold more through API pay - as - you - go billing. In a nutshell: ordinary capabilities will continue to be offered on a monthly basis, while high - end capabilities will be charged on demand.

Finally, do you subscribe to AI services in your daily use, and which AI do you use the most?

Original link: https://www.techspot.com/news/112759-openai-anthropic-cant-afford-have-everyone-use-ai.html

This article is from the WeChat official account "CSDN", author: Zheng Liyuan, reprinted by 36Kr with authorization.