Tokens & parameters - the two numbers that matter
A "token" is a chunk of text the model reads or writes - roughly ¾ of a word in English. "Parameters" are the model's internal dials - the things learned during training. More parameters generally means more capacity to capture patterns (and more cost/slowness). When people say "70B model" or "trillion parameter model," that's the dial count.
Text gets broken into tokens by a tokenizer before it ever reaches the model - "unbelievable" might become "un" + "believ" + "able." The model's context window is measured in tokens, not words or characters, which is why long documents "cost" more and can hit limits. Parameters are the weights in the neural network; they don't map to specific facts or concepts you could point to - knowledge is distributed across all of them.
Go to any tokenizer tool (search "OpenAI tokenizer" or use Anthropic's docs) and paste a paragraph of your own writing. Count how many tokens vs. words it produces. Then paste something in Arabic or another non-English language and compare the token count for the same amount of meaning - notice non-English text often costs more tokens.
What does a model actually read and write, under the hood?