The context may contain instructions, conversation, documents, multimodal input and generated output. A large limit lets more information fit but can increase latency, cost and memory demand.
A headline token count does not show how reliably the model uses distant details. Practical evaluation should test retrieval and reasoning across the claimed range with realistic projects and distractions.
ELI5
A context window is the amount of information an AI model can consider at one time. It can include the current request, earlier conversation, documents, instructions, and results returned by tools.
For example, if a long software project no longer fits, the system may keep the requirements and recent changes while summarizing older discussion. A larger window holds more, but filling it with irrelevant material can still confuse the task.
Does a one-million-token context window guarantee useful recall?
No. It describes capacity, while retrieval and reasoning tests show whether distant information is used reliably.
What costs can grow with context length?
Memory use, processing time and provider charges can increase as more tokens enter the request.
