Industry analysis firm SemiAnalysis has conducted a real-world test of the subscription plans of OpenAI and Anthropic. The results show that behind the seemingly affordable fixed monthly fees lies a huge computing power subsidy gap borne by the large model providers.

Testers purchased various subscription plans from both companies and continuously ran heavy tasks such as long-term programming and agents until they reached the weekly usage limits. They then calculated the theoretical costs based on the publicly disclosed API pricing, and the numbers were shocking.

Maxing Out Computing Power Subsidies

The calculations showed that if a user fully utilizes the OpenAI "ChatGPT Pro 20x" subscription priced at $200, the corresponding API billing could reach up to about $14,000. Similarly, the Anthropic "Claude Max 20x" plan, priced at the same level, could theoretically cost up to $8,000 in token expenses under extreme usage conditions.

This means that a small number of heavy users can consume the limited profit margin under the subscription model, leading the providers into serious losses. For ChatGPT Plus, a $20 entry-level subscription, OpenAI starts losing money on a user once their utilization exceeds 11.4%.

Enterprise-Level Task Distribution Becomes a New Trend

In this context, agent systems that rely on multi-step and autonomous tool calls are increasing cost pressures, with token consumption reaching thousands of times that of traditional single-turn conversations. Large companies including Microsoft, Meta, and Amazon have started to scale back previous practices of encouraging employees to extensively test AI in order to control rapidly growing internal bills.

To address the high computing costs, more enterprises are adopting a refined task distribution strategy, where complex problems are handled by top-tier models, while routine office work is delegated to cheaper or open-source models. This task routing method can cut overall AI costs by up to 95%, but it also forces large model providers to struggle to find a balance between user experience and high infrastructure investments.