See how much of a model’s context window your prompt actually uses, and which part of the conversation is consuming it.
Context utilisation
160,000 / 1,000,000
16.0%
80%95%
User
System 30,000(3.0%)User 90,000(9.0%)Assistant 40,000(4.0%)Free 840,000
Headroom
840,000 tokens free. Remember the model's reply comes out of the same budget — leave room for it.
Context is a ceiling, not a promise
Retrieval accuracy typically falls off well before the advertised maximum. A model quoting a 1M window will not reliably recall a detail buried at 800K.
llmtools is a vendor-neutral toolkit for working with large language models — token counting, cost estimation and model comparison, computed in your browser. OpenAI token counts are exact, using the same encoding the model does (o200k_base for GPT-4o, cl100k_base for GPT-4/3.5); Claude and Gemini counts are labelled approximations, since neither publishes a tokenizer that runs client-side. Model pricing is read from OpenRouter’s public catalogue once a day. Nothing you type is ever transmitted anywhere.