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

This article is an economic analysis of AI coding with a strong personal-experiment flavor. The author uses his own experience building an application of roughly 40,000 lines—still unfinished but functional—with Claude Code as an example, arguing that LLM coding genuinely lets people with an engineering background accomplish development work in a short time that would otherwise be very hard to achieve; but what he really cares about is cost. The author believes that simple conversations or low-risk tasks may now truly be cheap enough that it is not worth calculating carefully, but complex reasoning, code modifications, tool calls, and so-called "thinking" or recursive models generate large amounts of tokens the user never sees, including background attempts, retries, tool outputs, context re-sends, and verification workflows. Estimating with API prices, a complex multi-file coding task might fall in the range of tens of dollars; the author also mentions having seen a single query approach a million tokens, which, if calculated at output-token prices, could be quite expensive. The core inference of the article is that monthly plans like Claude Max, if fully used for high-intensity agentic coding, could have an actual API-equivalent cost that is several times the subscription fee—in extreme scenarios approaching or exceeding the subsidy magnitude of "paying 100 dollars while the provider bears over 1,000 dollars in cost." The author also places OpenAI/Codex in the same category of problem, arguing that both Anthropic and OpenAI may face similar pressure. It is worth noting that the article is not a formal financial report or vendor disclosure, but a mix of personal usage data, public pricing, model-assisted research, and an estimation framework; the author also explicitly acknowledges that the data is incomplete and the estimates are fragile. For Taiwanese developers and procurers, the point is not to accept the numbers wholesale, but to understand that subscription-based AI coding tools may not yet reflect their true unit economics, and that if quotas, prices, or model capabilities are adjusted in the future, team workflows and maintenance costs could all be affected.