Context window and output limit

The context window is counted in tokens. It holds the system prompt, the history, your documents and the question. On most APIs it must also leave room for the answer. When a request is longer than the window, the API rejects it or drops the oldest part.

A model’s maker states a window. A host may serve the same model with a shorter one, to fit more requests on its hardware. In the listings read for this site, Qwen3.8 27B appears with windows that differ many times over from one host to another.

So read the window in the host’s own row. A low token price on a short window is a different product from the same model at full length.

The output limit is the most a model may write in one response. It is set by the host, and for the same model it varies far more than the window does. A long report, a large code file or a long chain of reasoning can hit it and stop mid-sentence.

If your work produces long answers, compare this figure before price.

You pay for every token you put in the window, on every request. Filling a long window is a long input bill. A larger window is worth having when you need it. It is not a reason to send more than the task requires.

Where a host offers a cached input price, a long prompt that repeats costs less on later requests. Check for it when your prompts share a long opening.

Runcrate serves 124 models through one OpenAI-compatible API, each listed with its context length. You can try the same long prompt on several models from one key and one balance.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.