@rick.theengineer: Every time you talk to an AI, a meter is running 🚕 And it's not counting your words. It's counting tokens. Here's what actually shows up on your AI bill 👇 1️⃣ Tokens are small chunks of text. "Cat" is one token. "Unbelievable" gets split into three. Roughly, 1,000 tokens ≈ 750 English words. 2️⃣ You pay twice. Input is everything you send: your question, the documents you paste, and the hidden instructions behind the scenes. Output is everything the AI writes back. On Claude Sonnet 5.5 that's $2 per million tokens in and $10 per million out. Writing costs 5× more because the model reads your text all at once but writes one token at a time. 3️⃣ Trap one: thinking. Reasoning models think before they answer, and that hidden thinking is billed as output. A one-line answer can still cost a lot. 4️⃣ Trap two: no memory. Every new message resends the entire conversation. Message 50 carries all 49 before it. That's why long chats and AI agents get expensive fast. 5️⃣ The hero: caching. Save a big document once (a "cache write") and every reuse becomes a "cache read" at 90% off. In the example, 100 support-bot questions cost about $11 without caching and about $2 with it. Same answers, over 5× cheaper. 6️⃣ Batch. If you can wait for the answer, Anthropic and OpenAI both cut the price in half. 7️⃣ Pick the right model. GPT-6 Astra costs 100× more than GPT-6 Luna. The smartest model isn't always the right one for the job. Because with AI, you're not paying for answers. You're paying for tokens. 💬 Which one surprised you most: the thinking tokens, or the whole chat getting resent every time? 📌 Save this for the next time your API bill looks weird. (Prices as of October 2026, standard rates.) #ai #claude #openai #llm #promptcaching

RickTheEngineer

Open In TikTok:

Region: GB

Wednesday 07 October 2026 16:25:40 GMT