A model's context window is the number everyone checks. The number that actually determines your bill on a long request is a different one — the long-context repricing threshold, where one exists — and the two aren't the same figure, sometimes by a wide margin. This tool shows a real prompt against both, for the specific model you're planning around.

Two different ceilings, not one

The context window is a hard limit — the model can't accept more input than that, full stop. The long-context threshold, published for most but not every current model, is a soft one in the sense that crossing it doesn't fail the request; it reprices the entire thing at a higher rate, input and output both. For a model that publishes both figures, the threshold sits well inside the window, which means the practical ceiling anyone budgeting a bill actually cares about is usually the cheaper number, not the model's advertised maximum. A model with a very large window can still be a poor fit for a cost-sensitive workload if that workload routinely crosses the threshold long before it gets anywhere near the window itself.

Why this matters more for agentic sessions than single prompts

A one-shot prompt rarely gets close to either number. An agentic coding session accumulating file contents, tool outputs and prior turns into its own running context absolutely can — and it does so gradually, one turn at a time, which means there's rarely a single obvious moment where you decided to send a large request. The planner is built around that shape specifically: paste or describe what a session has accumulated so far, and see where it actually sits, rather than estimating from a single static prompt that doesn't reflect how the context actually grew.

What the tool shows

For the model you select, the planner shows your entered token count against the context window as a hard ceiling and, where the model publishes one, against the long-context threshold as a cost boundary — with an explicit flag the moment you cross the threshold, since that's the crossing most people miss until the bill looks larger than expected. Where a model doesn't publish the threshold at all, the planner says so plainly rather than assuming it inherits a sibling model's rule; several current models genuinely don't state one, and treating "not published" as "doesn't apply" and treating it as "definitely doesn't apply" are both guesses this tool refuses to make for you.

Reading the gap between window and threshold

The distance between a model's context window and its long-context threshold, where both are published, is itself useful information: a wide gap means you have real headroom to send a comfortably large prompt without triggering the higher rate, while a narrow one means the threshold is the binding constraint for almost any substantial request, and the window number on the model's spec sheet is close to irrelevant to what you'll actually pay. Different models in the same family can have meaningfully different gaps even when their windows are identical, because the threshold and the window aren't derived from the same thing.

Using this alongside the token-cost estimator

This planner answers "how close am I to a limit," not "what will this cost" — for the dollar figure once you know where a prompt actually sits, the token-cost estimator takes the same token counts and prices them against the model and service tier you're planning around, correctly applying the long-context multiplier the moment this planner tells you it's in play. Used together, one tells you whether you're about to cross a line and the other tells you what crossing it actually costs.

What to do once you're near the line

If a session is approaching the threshold and doesn't need everything it's currently carrying — old tool output that's no longer relevant, an earlier turn's full file dump that a summary would serve just as well — trimming before the next turn is a direct lever on which side of the line that turn lands on. That's a cheaper fix, in both effort and dollars, than discovering the crossing after the fact in a bill that jumped more than the token count seemed to justify — and it's a habit worth building into any long-running agentic loop, not just a one-time cleanup you do after the first surprising bill.

Context-Window Planner