The AI Cost Problem That Price Cuts Cannot Solve

DeepSeek has reduced the pricing for its flagship model 'V3-Pro' by 75%. While this may appear as good news for AI developers on the surface, there is no widespread sense among enterprises that 'costs have been reduced.' The reason is that the volume of tokens consumed by AI agents is growing at a faster rate than model unit prices are declining.

Over the past two decades, software infrastructure has followed a pattern where infrastructure costs fall annually while application performance continuously improves. AI was initially expected to follow the same trajectory, with developers assuming that as model performance improved and prices fell, inference costs would eventually reach negligible levels. However, the proliferation of AI agents appears to be undermining this assumption.

Traditional chatbots required only one model call for each user question. In contrast, when AI agents receive the same question, they execute multiple steps in sequence: planning, information retrieval, tool utilization, result verification, summarization, and additional judgment. While the user receives only one final answer, behind the scenes the model is called multiple times, incurring charges each time. The original text refers to this as the '100x problem.'

Looking at concrete numbers, the difference is extreme. In simple question-and-answer chatbots, the ratio of system processing charges to user input is typically around 1:5. However, in multi-step agents deployed for customer support or legal compliance tasks, this ratio can exceed 1:700. For example, even a single-sentence question like 'What inquiries did we receive from key customers last week?' can trigger seven billable operations—including repeated system prompt reads, search result retrieval, multiple model calls, and output formatting—consuming approximately 35,000 tokens in total processing.

This cost structure is also reflected in the pricing strategies of model providers. OpenAI's reported offer of $2 million in API usage credits to startups accepted into Y Combinator illustrates the cost level necessary for AI-native companies to survive their first year. This scale represents a stark contrast to earlier tech startups, which managed initial development with only thousands of dollars in cloud credits.

The fundamental issue is that reducing model unit prices does not reduce costs if the product architecture itself remains unchanged. Agents inherently transform a single user operation into dozens of billable operations. With each loop, prior conversation history, tool outputs, and reasoning progress accumulate and carry forward to the next step. Because nothing is omitted, token counts continue to mount. While price reductions may mitigate this trend, they do not resolve the structural problem.

The deeper AI agents are embedded in business systems, the more pronounced this cost amplification mechanism becomes. For developers and enterprises, model pricing alone is no longer the sole factor—the efficiency with which an agent's design uses tokens is becoming a critical metric directly linked to profitability. As model price competition intensifies, the focus is shifting to a new question: 'How can we achieve the same results with fewer tokens?' This represents a move toward architectural innovation as the next frontier.