This article explores the real-world costs of artificial intelligence, highlighting the significant energy and resource consumption required to train and operate AI models. It discusses the environmental and economic impact of the growing demand for AI infrastructure.
An AI feature in production incurred unexpectedly high costs due to longer customer prompts and added context retrieval, doubling token usage. The lack of cost attribution tools in the OpenAI dashboard makes it difficult to track which feature or model is driving expenses.
Token price is an easy but misleading metric for AI cost. The true measure should be the cost per finished task, as cheaper tokens might not actually be more economical if they require more work for the same usable result.
Claude Opus 4.7's new tokenizer is causing an effective 35% price increase for the same work due to higher token consumption compared to version 4.6. While reasoning improvements are real for complex tasks, the author plans to use 4.7 selectively and stick with 4.6 for tasks where token efficiency is key.
This article explores the author's experience of hitting the ChatGPT usage limit, revealing patterns of repetition and inefficiency in LLM interaction. It raises questions about hidden costs and the need for a perspective shift on how we use these tools.
The author observed a 40% increase in their AI bill (Claude) despite attempting to reduce usage. This unexpected cost was attributed to the habit of maintaining long, open conversations, which accumulated context and increased token consumption.
The article debunks the belief that AI consumes excessive water, stating a ChatGPT query uses little, but highlights the often-overlooked high cost of AI services. It details how the price of a ChatGPT Plus subscription represents a significant portion of income in various countries, especially outside the US.