Context visualiser
A context window is everything a model can look at in one request, and it is counted in tokens. Pick a size, then read what that much text actually looks like.
Pick a window
- 12,000tokens
- 8,783words
- 18pages of 500 words
- 47 KBas plain text
Measured on this sample, 1 token is about 0.73 of a word. Real text and other models give different ratios.
All five sizes side by side
Bars are drawn to scale. The largest is 256 times the smallest, which is why the first bar is only a sliver.
- 1K1,000 tokens, about 2 pages, 1× the 1K sample
- 4K4,000 tokens, about 6 pages, 4× the 1K sample
- 12K12,000 tokens, about 18 pages, 12× the 1K sample
- 64K64,000 tokens, about 95 pages, 64× the 1K sample
- 256K256,000 tokens, about 375 pages, 256× the 1K sample
The 12K sample, zoomed out
Loading the 12K sample (47 KB)…
Read all of it
This is the complete 12K sample, not a preview. Scroll it, search it with your browser, or download the file above. The words are invented: a made-up field log with no meaning beyond filling the window.
Loading the 12K sample (47 KB)…
Everything shares one window
The window is not just your question. Five things compete for the same space in one request.
- InstructionsThe system prompt, rules and examples that set up the job.
- Your promptThe question or task in front of the model right now.
- ConversationEarlier turns, both yours and the model’s. Long chats fill the window quietly.
- Tool resultsWhatever searches, files and code runs hand back. These are often the biggest item.
- Generated outputThe reply. It uses the window as it is written, and joins the conversation for the next turn.
Try a budget
Set what goes in and how much reply you want. The window is the 12K size you picked above (12,000 tokens). Push a number too high to see an overflow.
| Instructions | 600 | 5.0% |
|---|---|---|
| Your prompt | 200 | 1.7% |
| Conversation so far | 2,500 | 20.8% |
| Tool results | 4,000 | 33.3% |
| Output you reserve | 2,000 | 16.7% |
| Needed in total | 9,300 | 77.5% |
What a big number does not promise
Max output is its own limit
The window is the total. Providers also cap each reply, and that cap can be far lower than the window. A 256K window does not mean a 256K answer, and no model here is promised to produce one.
Hidden reasoning varies
Some providers spend tokens on reasoning you never see, and treat those tokens differently. Check how yours counts them against output and the window.
More context is not perfect recall
A model can miss or muddle details buried in long input. A bigger window means more fits, not that all of it is used well. Test with your own material before you rely on it.
Token counts depend on the model
Counts here use cl100k_base, an OpenAI tokenizer, through js-tiktoken 1.0.21. Another model can split the same text into more or fewer tokens, so treat these numbers as a yardstick, not a quote.
The samples are exact
Each file was generated with a fixed seed, then counted to land on exactly 1,000, 4,000, 12,000, 64,000 and 256,000 tokens. The project’s tests count them again.
It all stays in your browser
No model is called. Budget calculations and the simulation run locally, without uploading your settings. The samples are plain text files served with this page.