Context limits and token arithmetic

How to estimate a token count from something you can see, what fits in each common window, and how the window behaves as it fills.

Tokens per unit of text

English prose	about 1.3 tokens per word · about 4 characters per token	So 750 words is roughly 1,000 tokens. Accurate to within about 10% across ordinary English for every widely used tokenizer.
Code	about 2-3 characters per token	Denser than prose because identifiers, punctuation and indentation all split. A line of code is roughly 8-15 tokens; a 1,000-line file is 10,000-15,000.
Languages other than English	1.5× to 3× more tokens for the same meaning	Worst for non-Latin scripts and for languages with rich morphology. The same message can cost three times as much in Hindi as in English, which is a real product and pricing issue.
JSON and structured data	roughly 2× the token count of the same content as prose	Keys repeat on every record and punctuation tokenises separately. For bulk extraction, a tabular format costs materially fewer tokens than nested JSON.
Numbers	often one token per one to three digits	Which is part of why models are unreliable at arithmetic on long numbers: the digits are not individually addressable in the way the task requires.

English	characters ÷ 4	Or words × 1.33. Either is close enough for budgeting; use the real tokenizer when you are near a hard limit.
Code	characters ÷ 2	Conservative, which is the right direction to be wrong in when you are checking whether something fits.
A page of prose	about 500 words · about 650 tokens	A4 or US Letter, single spaced, ordinary margins.
1 MB of plain text	about 250,000 tokens	Useful for sizing a corpus before you start: a 40 MB document set is roughly 10 million tokens to embed.
An hour of speech	about 9,000 words · about 12,000 tokens	At roughly 150 words per minute of natural speech. A meeting transcript is smaller than people expect.

8k tokens	about 6,000 words · roughly 12 pages	Enough for a long document or a short conversation, not both.
32k tokens	about 24,000 words · roughly 50 pages	A long report, or a RAG prompt with generous retrieved context and a working conversation.
128k tokens	about 96,000 words · roughly 200 pages	A short novel, or a medium codebase. Common ceiling across current models.
200k tokens	about 150,000 words	Roughly two full-length books, or a codebase of a few hundred files.
1M tokens	about 750,000 words	Available on some models. Cost and latency scale with what you actually send, so a large window is a capability, not an instruction to fill it.
Output limits are separate	typically a few thousand to tens of thousands of tokens, well below the input window	And on most APIs input and output both count against the context limit, so the reserved output budget reduces what you can send.

Lost in the middle	recall is best at the start and end of a long context, worst in the middle	Measured repeatedly across models. Put the most important material at the top or the bottom, and do not assume a large window means uniform attention.
Prefill cost	attention work grows with the square of sequence length	So doubling the prompt more than doubles time to first token. The linear term is the weight loading; the quadratic term takes over at long context.
KV cache cost	grows linearly with context, and it is per concurrent sequence	This is what caps concurrency on a self-hosted service. Long contexts and high concurrency compete for the same memory.
Cache-friendly ordering	stable content first, varying content last	Prefix caching only matches from the beginning. One changed token near the top invalidates everything after it, so a timestamp in the system prompt can cost you the whole cache.
Effective versus advertised context	usable quality often falls well short of the stated limit	Especially for models where long context was extended after pretraining. Test on your own longest realistic inputs rather than trusting the number on the model card.

Figures last checked 2026-09-13. Several of these age; check anything you are about to rely on.

Where these numbers come from