Why would I deliberately fill a context window?

Long-context models often degrade well before their hard limit — recall drops, latency climbs, and cost scales with input tokens. Stress testing at 50%, 75%, and 95% fill reveals where your specific workload starts to suffer.

How accurate is the token count?

It uses an empirical characters-per-token ratio that differs by content type (prose, code, JSON). It is an estimate within a few percent, which is close enough for stress testing. For billing-exact counts use the provider's tokenizer.

Does the filler get sent anywhere?

No. The text is generated entirely in your browser. You copy it and paste it into your own LLM call, so nothing touches our servers.

What is a needle-in-a-haystack test?

You hide a specific fact (the needle) inside a large filler context (the haystack) and ask the model to retrieve it. This tool builds the haystack at a precise fill level so you can run that test repeatedly at different loads.

Why does filler type matter?

Code and JSON tokenize more densely than prose because of punctuation and symbols, so a 1,000-token JSON block is far shorter in characters than 1,000 tokens of lorem. Matching filler type to your real data makes the test representative.

Context Window Stress Tester

Context window stress tester

Modern models advertise huge context windows — 128K, 200K, even a million tokens — but capacity is not the same as quality. Recall accuracy, latency, and per-call cost all change as you fill the window. The context window stress tester generates filler text sized to land on an exact percentage of a model’s window so you can probe that behaviour deliberately instead of guessing.

How it works

You pick a model (or enter a custom window size), a target fill percentage, and a filler type. The tool converts your target percentage into a target token count, then generates that many tokens of filler using a per-type characters-per-token ratio: prose is roughly 4 characters per token, source code about 3.2, and JSON about 2.8 because of dense punctuation. It builds the text in the browser and reports the achieved token estimate and character count so you can paste it straight into a real call.

Tips and notes

Test the curve, not one point. Run 50%, 75%, and 95% — degradation is rarely linear.
Match filler to your workload. If you feed the model JSON, test with JSON filler so token density is realistic.
Embed a needle. Drop a unique fact near the start, middle, and end of the filler, then ask the model to retrieve each to map positional recall.
Watch cost. Input tokens are billed every call; a 95%-full 200K window is expensive to probe repeatedly.