Will your prompt fit? Check context windows across 20+ frontier LLMs.

How this checker works

Three numbers decide whether your text fits a model: how many tokens it becomes, how large that model's context window is, and how much of the window you need to leave free for the reply. The first is the only one that requires real work, because there is no universal answer — the same string becomes a different number of tokens on every model family.

Token counts: one family exact, the rest calibrated estimates

For OpenAI models we run the real thing. The o200k_base tokenizer is loaded in your browser and your text is encoded with it, so the count is not an approximation — it is what the API will bill you for. Anthropic, Google, and Meta do not publish a client-side tokenizer, so for those families we estimate from a characters-per-token ratio measured on mixed English and code:

Ratios last calibrated 2026-05-07. Estimated rows are badged in the results table so you always know which of the two you are looking at. Treat an estimate as accurate to within a few percent on prose, and as less reliable on dense code, JSON, or non-Latin scripts, all of which tokenize more expensively than English.

Why the verdict is not simply “fits” or “doesn't”

The model writes its answer into the same window your input occupies, so a prompt that technically fits can still fail the moment the reply runs long. We score against input plus a reserved response budget — the slider in the tool — and report three states rather than two:

• Fits — under 50% of the window used. Room for a long reply and a few follow-up turns.
• Tight — 50% to 95%. It will go through, but a multi-turn conversation will not, and quality on some models degrades well before the hard limit.
• Over — 95% or more. Expect truncation or an API error. Split the input or move to a larger window.

How the cost figure is built

Cost is input tokens at the model's input rate plus your reserved response budget at its output rate, both per million tokens, for a single call. Two deliberate omissions: we do not apply prompt-caching discounts and we do not apply batch pricing, because both depend on how you call the API rather than on what you are sending. The number you see is therefore the conservative one — the ceiling for one uncached, non-batched request. If you cache a long system prompt across many calls, your real cost will be meaningfully lower.

Window sizes, output limits, and prices come from the same hand-maintained catalog behind the LLM picker, checked against each provider's own pricing and model documentation. Last verified 2026-08-16. Providers change prices without notice, so if you spot a stale figure, tell us — those corrections ship as code changes, usually within a day.

Your text never leaves your browser

Tokenization, fit scoring, and the cost arithmetic all run on your device. Files are parsed locally too — PDFs and Word documents are read in the browser, never uploaded. Nothing you paste is sent to us, logged, or stored, which is the reason this tool is safe to point at a contract or an internal document. Shared links encode the model and response budget you picked, never the text itself.

What this tool will not tell you

Fitting is a floor, not a goal. Several models retrieve noticeably less well from the middle of a very long context than from its beginning or end, so a document that fits at 90% of the window may still produce worse answers than the same document summarised down to 30%. Nor does a fit tell you the model is the right one for the job — for that, compare candidates in the LLM picker. And on multimodal input, images and audio consume tokens by a formula of their own; this checker only measures text.