Tokens & Context Windows Cheat Sheet

What counts, what it costs, and what falls out of the window

Tokens are the unit of both billing and memory. Estimating them well is most of cost control and all of context management.

Category: AI Models Difficulty: Intermediate Version: 1.0 Updated: August 4, 2026 Author: Sabir

Estimating tokens

Content	Rough tokens	Note
English prose	~1 token per 4 characters	About 0.75 tokens per word
Code	~1 token per 3 characters	Punctuation and indentation are dense
JSON	Higher still	Braces, quotes and keys all cost
Non-Latin scripts	2–3× more per character	Budget generously for CJK and Arabic
An A4 page of text	~500–700 tokens	Useful for sizing document pipelines

What occupies the window

The system prompt — every single turn — It is re-sent with each request, not remembered
The full message history you choose to send — The API is stateless; you own the memory
Tool and function definitions — Their JSON schemas are prompt text
Tool results fed back in — Often the largest single contributor
The reply being generated — Input and output share the window

Truncation strategies

Strategy	Keeps	Loses
Sliding window	The last N turns	Anything agreed early on
Summarise-and-drop	A rolling summary plus recent turns	Detail, gradually
Pinned + recent	Explicitly marked turns, plus recent	Needs someone to choose the pins
Retrieval over history	Whatever matches the current turn	Continuity between turns

FAQs

Does a bigger context window mean I can stop managing context?

No. Cost scales with what you send, latency grows with it, and retrieval accuracy over a very long context is not uniform. A big window buys headroom, not a free pass.

Why does my token count differ from my word count?

Tokens are sub-word pieces. English prose runs about 0.75 tokens per word; code, JSON and non-Latin scripts run considerably higher because punctuation and rare characters split into more pieces.