Context window checker: will your text fit?

A model’s context window is the most tokens it can handle in one request, including its own reply. Windows range from a few thousand tokens to 2,000,000 for Grok 4.20, the largest in our data. The checker adds up everything a request carries and compares it with each model’s published limit. For background, see context windows explained.

Does the reply count toward the context window?

Yes, and the reply is only one of several things that share the window. A request is more than the text you paste. It also carries the system prompt, any tool or function definitions, the earlier messages in a conversation, and room for the answer. The checker splits these into three parts: your text, other prompt tokens, and the reply. The bar for each model shows all three against its window, and a line marks the edge when the request goes past it.

For example, a 200K-token window might hold system prompt (2,000), tool definitions (3,000), earlier messages (15,000), attached document (150,000), room for the reply (8,000): 178,000 tokens in all, leaving 22,000 spare. In the checker, the document is your text, the first three go in “other prompt tokens”, and the last is the reply room.

How are the tokens counted for each model?

Different models split the same text into different numbers of tokens, so the checker counts per model. For OpenAI models and for open-weight models whose tokenizers are published (such as DeepSeek, Qwen and Mistral), it runs the model’s own tokenizer in your browser, the same engine the token counter uses to count tokens per model, and the count is exact.

Claude, Gemini and many other models don’t publish their tokenizers. Rather than pretend to know their count, the checker gives a range. We measured the published modern tokenizers against OpenAI’s on more than 1,500 texts in many languages: they used between 0.61 and 2.0 times as many tokens, and most stayed within 0.88 to 1.43 times. So the range runs from 0.5× to 2× OpenAI’s count, widened to 3× when most of the letters are in a non-Latin script, where tokenizers differ most. A model is only marked “Fits” if the top of the range fits, and only “Too large” if the bottom of the range doesn’t. In between, it says “Too close to call”, and the official count (free, with your own API key) gives the exact answer.

What happens when a prompt is too long?

Most APIs reject it with an error, so you have three main options. Choose a model with a larger window: the checker lists the cheapest current models with room for your request. Split the text into parts that each fit, and process them one at a time; the checker tells you how many parts a given model needs, and you can split a long document into chunks. Or send less: summarise long documents first, or retrieve only the relevant passages instead of the whole text. The last option is often the best, because very long prompts cost more and run slower.

A window that fits isn’t the same as a window the model uses well. Research on long inputs found that models recall information at the start and end of a prompt better than information “lost in the middle”, so put the most important instructions and passages near the start or end.

Context windows and reply limits come from the OpenRouter models API and the LiteLLM model list, refreshed daily (last updated 2026-10-16). Some providers offer larger windows on request or on special tiers; the checker uses the standard published limit.