One shared budget

The context window holds input tokens AND output tokens together. Every token the model reads and every token it writes spends the same budget.

A long answer leaves less room for input on the next turn. Budgeting must reserve space for the reply, not just the prompt.

What spends the budget

Five contributors spend the budget on a typical assisted task. All five are input except the generated reply.

Estimating before pasting

Token counts are tokenizer-dependent, so estimation uses approximate EXAMPLE rates. A common rough rule is about 3 to 4 characters per token for English prose with common tokenizers, and fewer for code, numbers, hex and non-Latin text.

The calc block plans an EXAMPLE session with 4 chars per token for prose and 3 for code. Real counts vary by model and tokenizer: measure them on the tool in use.

1
2
3
4
5
6
7

EXAMPLE rates (stated parameters; real tokenizers vary):
prose 4 chars per token, code 3 chars per token
system prompt: 400 chars prose -> 400 / 4 = 100 tokens
recent chat: 3000 chars prose -> 3000 / 4 = 750 tokens
spec extract: 2400 chars code -> 2400 / 3 = 800 tokens
answer reserve: 2000 chars prose -> 2000 / 4 = 500 tokens
planned total = 100 + 750 + 800 + 500 = 2150 tokens

Common mistakes

Related topics

• When the window fills

• Tokens and tokenization

• Cost arithmetic