Most people using AI tools have no real idea what a token is, how many they're using, or why the same text costs differently depending on the model. This episode is the missing primer: what tokens actually are, why text doesn't break up the way you'd expect, how the billing works, and seven common misconceptions that reliably confuse people who should know better by now.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "LLM Tokens Masterclass: Curriculum-Grade Research" (Dr. Priya Nair, June 12, 2026). Pricing data retrieved live from official Anthropic, OpenAI, and Google documentation.

- What a token is — and why rare, unusual, or technical words cost more than common ones

- BPE: the compression algorithm that trained your model's vocabulary from scratch

- Input vs. output: why what you type costs a fraction of what the model writes back — and the exact math behind the gap

- How billing works: per-million pricing across Claude, GPT, and Gemini, with live figures

- Context windows as working memory — what it means to fill one, and why "context rot" sets in before you hit the hard limit

- Practical rules of thumb: 750 words is roughly 1,000 tokens, and what breaks that rule

- Seven misconceptions corrected: tokens aren't words, context isn't memory, more context isn't always better output

- The kicker: Anthropic changed their tokenizer starting with Opus 4.7 — the same text now produces up to 35% more tokens than it did on older Claude models