The way humans naturally type—typos, shorthand, filler words—directly impacts how much they pay for AI API calls. A developer discovered this when a simple five-word prompt with two spelling mistakes cost 13 tokens, but only 6 tokens after corrections. Common typing habits like swapped letters and dropped characters fragment differently in tokenizers than clean text.

OpenAI's tokenizer and Anthropic's Claude produce different counts on identical input—Claude generally generates more tokens. Shorthand often backfires: "pls" costs more than "please" on Claude, "thx" exceeds "thanks," and "w/o" balloons to 3 tokens versus "without" at 1. Technical identifiers like UUIDs consume significant budget—a single UUID can reach 24-26 tokens, while timestamps and file paths add up quickly.

Conversational fillers like "just," "basically," and "actually" create unnecessary overhead, as do expressive patterns like "Good..." or "Yes!!" that add tokens without improving results. The model may recover meaning from all of this. Billing does not. Humans optimize for keystrokes; tokenizers optimize for common text—and those priorities don't align.