Sources checked
Budget instructions, evidence, history, and output
For an original support workflow, count the system instructions, recent messages, and relevant policy passages before adding more history. Keep space for the answer and inspect whether the application truncates older material.
Input capacity does not guarantee evidence use
A larger context limit does not guarantee that every detail will be used reliably. Test whether the model finds the relevant evidence in realistic long inputs rather than only checking that the request is accepted.
What the limit includes
The window counts tokens, not pages. It usually has to hold the instructions, the conversation so far, any retrieved passages, tool results, and the answer the model is about to write.
If the answer needs room, you cannot fill the entire window with source text. Some interfaces hide system instructions that still consume tokens.
Providers differ on what happens at the edge. One truncates the oldest turns.
Another rejects the request. A third silently drops a middle section.
Read the behavior for the endpoint you call. A demo that pasted a short article does not tell you what happens to a 200-page file.
Why a long prompt can still miss the fact
Research on long inputs has found that models often use the beginning and the end of a prompt more reliably than the middle.
That result is about those tests, not a law of every new model, but it is a reason to test placement.
Put the clause that must be obeyed, and the passage that must be quoted, in a position you have checked.
Retrieval is often a better fit than pasting everything. Select a few passages that a search step can justify, and keep the question next to them.
If you do send a long file, score the model on questions whose answers sit at the start, the middle, and the end. A window size on a pricing page is a capacity number, not a comprehension score.
- Reserve tokens for the answer and for tool results.
- Confirm whether overflow truncates, errors, or drops the middle.
- Test facts that appear at the start, middle, and end.
- Prefer a retrieved passage over an untested dump of every file.
What to record before you rely on it
Write the window size in tokens, what the provider does at the limit, and how many tokens your instructions already consume before the user types.
Test one fact at the start of a long document, one in the middle, and one at the end. Save those three results with the model id.
If a new model claims a larger window, rerun the same three probes. Capacity on a pricing page is not a substitute for that note.
Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test.
Common questions
If the window fits my document, will the model use all of it?
Not necessarily. Fitting is a limit on size.
Use is a behavior you have to test, especially for details that sit in the middle of a long prompt.
Do chat history and the new question share one window?
Yes. Earlier turns, hidden instructions, and the new question all spend the same budget.
A long chat can push the original document out even though the window looked large at the start.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





