Token: why the AI ​​bill rises without anyone using it anymore

Оригинальный автор：Sem Atalho

Первоисточник：https://www.youtube.com/embed/oTfEqF0qOuM

The signature is the same. The team is the same. And the AI ​​bill came up bigger again. It's not a billing bug or vendor issue. That's how it works and to understand it you just need to understand one word: token. Token is the piece of text that the model collects. It's not a word, it's not a letter. It's a piece. And the part that explains your bill is this: with each new message, the entire conversation goes together again. He is the lawyer who re-reads the entire file before writing each petition and charges for reading it. In this video I open the terminal with a token counter in view and show: • a simple word turning into several tokens • why the first call goes out with 49 thousand tokens sent, even with the conversation being "empty" • Portuguese costing more tokens than English to say the same thing • the same JavaScript function, with and without spaces and tabs: 124 tokens difference • why saving money on the way there matters less than it seems what costs dearly is the way back And I close with the three consequences for those who respond for delivery, and with the phrase that changed how I look at measurement: metrics shape behavior. 📄 THE PAPER I CITE "The Hidden Cost of Readability: How Code Formatting Silently Consumes Your LLM Budget" — Pan, Sun, Zhang, Lo and Du, accepted at ICSE 2026 https://arxiv.org/abs/2508.13666 ⚠️ CORRECTION: in the video I talk about "40% savings". The paper's number is 24.5% average reduction in input tokens, measured in code (Java, Python, C++ and C#, ten models), with no loss of performance. The 36% that appears in the study is something else: reducing the size of the output code. Here's the record — the link is there to check it out. 🎯 FOR THOSE WHO ARE DELIVERY MANAGER, TECHNOLOGY LEADERS, PO, FACTORY MANAGER. Who accounts for the AI, not just for using it. CHAPTERS 0:00 The signature is the same and the bill is bigger 0:28 What is a token — it's not a word, it's not a letter 1:14 The lawyer's analogy: why everything is resent 2:30 In the terminal: "development" becoming several tokens 3:10 Why 49 thousand tokens are already sent: the context is never empty 4:10 Question and answer: everything becomes a token 4:47 Portuguese costs more token than English 5:42 Formatted code vs. code without spaces and tabs 6:43 The paper — and why saving money on the way is not the point 7:28 In Python, indentation is syntax 7:58 Three consequences for those responsible for delivery 8:30 Measure by token? Why PR? The question that matters 9:05 Metrics shape behavior 9:55 Next episode: what is context 💬 Do you know how much an AI demand from your team costs? Not the subscription: the demand. Tell me in the comments if this number exists there. I answer everyone while the channel is small. ABOUT THE CHANNEL Sem Atalho is AI for those responsible for delivery. I use the channel as a study tool: I apply the Feynman Technique, which says that explaining without jargon is proof that you understand. There is no shortcut — this is the method and this is the name. Series Without Jargon, one word per episode. Next: Context. @ia_sem_atalho #InteligenciaArtificial #LLM #CustoDeIA #EngenhariaDeSoftware