Tokenization

Tokenization is discussed on ToTheMoon in two senses. The first is technical: how speech and text are cut into tokens a model can process. The second, far more frequent, is economic: tokens as the unit of spend, limits and the right of access to strong models. The central question is posed in episode 139 — what working with the best models really costs, and whether access to them will become a privilege.

The technical side came up with ChatGPT's new voice: the model works not with a raw audio wave but with audio tokens produced by a neural codec, much like chunks of text (episode 099); episode 104 explains why the API is priced by token size and why a reasoning model is slow. Early in 2026 the economic angle appears: in Codex you are burning tokens and the supply is limited (episode 096), and in a discussion of Musk's remarks on chips the hosts note the position that whoever sells more tokens, and can hand them out cheaply, wins (episode 103).

By summer 2026 tokens have become personal experience. Access to Fable paid by the token is called access only for people with very large budgets (episode 121). Meta went through seventy-three trillion tokens in a month, and the hosts object that burning tokens is justified only when it produces product experience and data (episode 124). In episode 130 tokens in Codex and Claude are consumed differently and sometimes vanish almost instantly, so the real comparison is how much of a project gets finished. In episode 139 a personal counter showed forty billion tokens, and the conclusion is blunt: heavy spend does not guarantee a result.