Context Window Math — Tokens, Order, and What Stays

Tokens are the unit, not characters

Every model bills, limits, and reasons in tokens. A token is roughly ¾ of an English word — but it varies by language (Korean is dense, code is dense in different ways) and by tokenizer. Anthropic, OpenAI, and Google all use different tokenizers; the same English paragraph occupies a different number of tokens in each. Pretend they're the same and your context-window math will be off by 20-30%.

Three numbers to know about your model

Total context window — input + output. Claude Opus 4.7 has a 1M-token mode; OpenAI GPT-5.5 and Gemini 2.5 Pro are at 200k–2M depending on tier.
Output cap — practical max for a single response, often much smaller than the context window.
Effective attention budget — past a certain length, attention degrades. The model can technically read 1M tokens, but it doesn't attend to the middle equally well. Long-context limits is a real concern, covered in lesson 7.

Order is part of the prompt

Models pay disproportionate attention to the start and end of the context — the lost-in-the-middle effect. Put your instruction at the top, the most important evidence at the top of the document block, and a short reminder of the instruction at the bottom for very long contexts.

Code

Counting tokens for a prompt·python

import anthropic

client = anthropic.Anthropic()
resp = client.messages.count_tokens(
    model="claude-opus-4-7",
    system=open("prompts/agent.md").read(),
    messages=[{"role": "user", "content": user_text}]
)
print(resp.input_tokens, "input tokens")

Sandwich the instruction in long contexts·markdown

# Top of context — full instruction
[5-clause contract]

# Documents
[hundreds of thousands of tokens]

# Bottom of context — short reminder
Return JSON matching the schema above. Cite each claim with a doc id.

Context Window Math — Tokens, Order, and What Stays

Tokens are the unit, not characters

Three numbers to know about your model

Order is part of the prompt

Code

External links

Exercise

Progress

Comments 0