Skip to content
Back to full roadmap
topicfoundation

Tokenization

Token = the atomic unit the model sees. Token count = cost + context consumption.

2 hours2 hours3 resources1 prereqs

Models don't see sentences, they see tokens. For GPT/Claude, 1 token ≈ 4 English characters but only ≈ 1.5-2 Turkish characters. Turkish costs more because BPE was trained heavily on English.

Practical:

  • Same content in EN may use 30-50% fewer tokens
  • Stripping JSON whitespace saves money
  • Long-context calls are expensive; send only what's needed

What you'll gain

You'll estimate token counts at a glance and budget cost intentionally.

Prerequisites

Resources(3)