Context Window Guide

1. The short answer

A context window is the total number of tokens a model can attend to in one request — your input plus its own output. It is the one number that decides whether a task fits at all: summarise a 300-page document and the window is the whole ballgame; draft a two-line reply and it is irrelevant.

Bigger is not automatically better. A large window is generally slower and is priced per token, so a long prompt costs more on every single call. If you are skimming: for everyday work 128K is comfortable, and only whole-document tasks need the million-token tiers.

2. Context window and price, side by side

Across the 145 text models on Models today, the advertised windows cluster into three bands. Representative models from each, with the rates that actually decide your bill:

128K – 400K — 72 models

1M+ — 57 models

2M – 20M — 9 models

3. What one 1M-token prompt costs

Rates are per million tokens, so the arithmetic is mostly a question of which band you are in. Input only, one call:

Two things this table is really telling you. First, a million-token prompt is cheap on a budget model and expensive on a flagship — a factor of roughly a hundred between the rows above. Second, output is priced separately and is usually several times the input rate, so a long answer costs more than a long question.

4. Four moves that cut the bill

1. Summarise and chunk. Long document in, short summary per section out, then a final pass over the summaries. More calls, far less context per call.
2. Retrieve instead of stuffing. Send the three paragraphs that matter rather than the whole corpus.
3. Trim the history. Long chats grow monotonically. Drop or summarise turns older than the last few.
4. Move up a tier only when the task needs it. If a task genuinely requires the whole thing in view at once, pay for the window — and check the input rate first, because that is what a long prompt multiplies.

5. Is “1M” a marketing number?

No — where it is offered, 1M is a real ceiling; the marketing is in implying that all of it is available for input.

Three tests, in order of how often they catch people out.

• Read the input rate, not the window. A headline window with a cheap input rate is genuinely useful. A headline window at flagship rates means every long prompt is a flagship prompt.
• Check whether the output budget is separate. Some models reserve part of the window for reasoning or cap the output independently, so the usable input is smaller than the advertised number.
• Test retrieval quality at length. A model that accepts 1M tokens does not necessarily find one fact equally well at token 10 and token 900,000. Run your own documents through it before you build on it.

Band sizes on this catalogue are a decent reality check: the 1M tier holds the most models, while only a handful go beyond 2M. If your task is not in that handful, the bigger window is not what you are buying.

6. What AICraft does about it

• Every model row shows its window on Models, so you can filter by it without leaving the page.
• Model IDs pass through unchanged. Name a model explicitly and that is the model you get — the window you picked is the window you get.
• The Auto Router groups by task, not by window. Sending "model": "auto" picks a chain for the task it detects. If your request depends on a specific window, name the model instead of using auto.

Note and disclaimer

Models, context windows and rates change often. Every number on this page is read from the live catalogue — Models for availability and context windows, Pricing for the billing table — and the live pages are the source of truth wherever this page and they disagree.

Context figures are the maximum input window published for each model. Rates are US dollars per million tokens. Only representative models are listed here; the full catalogue is on the Models page.