PickWorth

Method

Context windows, explained without the headline numbers

Every vendor now advertises a very large number. The number is real and mostly beside the point. What matters is how the window behaves when it is full, and whether your work ever fills it.

22 October 2026 PickWorth Editorial

The short answer. A context window is the model's working memory for a single conversation — everything it can hold at once. A larger window is genuinely useful for two jobs: reading a lot at once, and long agentic work. The two questions that decide whether a tool's window matters for your work: how long is the input you are actually feeding, and how much of it does the model have to hold while it acts?

What is a context window?
The model's working memory for a single conversation — everything it can hold at once: your prompt, the conversation so far, and the tools and files it is acting on.
Does a bigger window automatically mean better?
No. A bigger window is only useful for two jobs: reading a lot at once, and long agentic work. For short tasks, the window is not the constraint.
What are the two questions that decide whether the window matters?
How long is the input you are actually feeding, and how much of it does the model have to hold while it acts? If the answer is short, the window is not the constraint.
Why does a big window sometimes feel slow or inconsistent?
Longer input means more to hold, more to attend to, and more to compute: the quality-per-token drops as the window fills. This is why a big window is not automatically better.
What is the practical takeaway?
Buy the window for the work you actually do, not the spec you are impressed by. Read more at once, or run longer agentic loops — that is when the bigger window pays.
Does PickWorth take money from any vendor?
No. Reviews are written by a single editorial team; the model docs and context limits are verified from each vendor's official pages.

What a context window is

A language model has no memory between conversations. Everything it knows about what you are doing right now — your question, the document you pasted, the answer it gave three turns ago, the system instructions — all of it lives in a single buffer for the duration of that conversation. That buffer is the context window, measured in tokens, which are roughly three-quarters of a word each. When the conversation exceeds the window, something has to go: either the oldest content is dropped or compressed, or the tool refuses to continue. There is no third option, and no model is exempt.

Why bigger is not automatically better

Two reasons, and the second is the one the spec sheets leave out. First, cost: a bigger window does not change the price per token, but filling it does — a tool that spends its whole window on every request is an expensive tool even at a low rate. Second, and more important, attention. Holding a million tokens is not the same as using a million tokens well. Models degrade on very long inputs: the material in the middle of a long document is recalled less reliably than the material at the start and the end, and an instruction buried under forty pages of context may be followed less faithfully than the same instruction on its own. This is why two tools with the same advertised window can give visibly different answers on the same long input — and why a review that reports only the spec sheet number is reporting the least interesting fact about the window.

The two questions that decide it

Does your work ever fill the window? If your typical input is a question and a paragraph, the window size is irrelevant to you — any current tool is deep enough, and the differentiator is somewhere else entirely. If your typical input is a full report, a long transcript, or a codebase, the window is the first number that matters. And: what happens when it fills? Some tools compress gracefully, keeping the thread of a long conversation coherent. Others start losing the beginning, which shows up as a tool that contradicts something it agreed to an hour ago. This behaviour is what you are really buying, and it is almost never on the pricing page.

What to actually do about it

Test it yourself, and test it in a way that fails. Paste in a document longer than you expect to need, then ask a question whose answer is in the middle of it. Then ask a question whose answer contradicts something at the start. The tool that gets both right is using its window; the tool that gets the second one wrong is holding the text but not reading it. That test takes five minutes and tells you more than any benchmark, because it measures the thing you will actually notice at nine in the evening on a long job.

Related