vMira

What is Context window?

The context window is the maximum number of tokens an LLM can consider in a single request — including the system prompt, conversation history, retrieved documents, and the user's current question. Anything beyond the limit must be truncated, summarized, or otherwise handled before being sent to the model.

Also known as: context length, token window, input window

Why it's measured in tokens, not words

LLMs operate on tokens, which are subword units produced by the model's tokenizer. A token can be a whole short word, a fragment of a long word, a single character, or a punctuation mark. As a rough rule, English text is about 1 token per 0.75 words (so 1000 tokens ≈ 750 words). Other languages — especially CJK and non-Latin scripts — tokenize less efficiently, sometimes 2-4 tokens per word.

Context window sizes in 2026

Common ranges: 4K-32K tokens for small/cheap models, 128K for mainstream models, 200K-1M for premium tiers. vMira's Pro tier supports 1M-token context. Larger context windows let you paste entire codebases, long PDFs, or many documents at once without RAG-style chunking.

Larger isn't always better

A long context window doesn't mean the model attends equally to every token. Recent research (the "lost in the middle" effect) shows accuracy on retrieval-style tasks drops sharply for information in the middle of a long context. Models are usually best at finding facts near the start or end. Putting the most important context first or last, or using RAG to feed in only relevant chunks, often beats stuffing everything into a giant prompt.

Cost implications

Input tokens are typically cheaper than output tokens, but a 100K-token input request still costs 50× a 2K-token request. Long-context workflows can quickly become expensive. Strategies to manage cost: summarize earlier conversation turns, use RAG for selective retrieval, cache the static portion of the prompt (prompt caching, supported by most major providers as of 2026), and pick the smallest model that handles the task.

Last updated · First published

Related terms

Large language model (LLM)

A large language model is a neural network trained on a massive corpus of text to predict the next token given the previous ones. Modern LLMs have billions to trillions of parameters and can generate coherent prose, answer questions, write code, and perform reasoning tasks well beyond simple text completion.

Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation (RAG) is a technique where an LLM, instead of relying solely on its training data, retrieves relevant text from an external source (documents, a vector database, the web) at query time and includes those retrieved passages in its prompt before generating an answer.

Thinking mode

Thinking mode is an LLM operating mode where the model spends extra compute on internal reasoning steps before producing a final answer. It significantly improves accuracy on math, science, logic, and analytical tasks at the cost of higher latency and price per response.

Embedding

An embedding is a fixed-length vector of floating-point numbers that represents the semantic meaning of a piece of text (or image, audio, etc.). Two embeddings are similar (small cosine distance) when their source texts are semantically similar — even if the words differ.

Try Context window in vMira

Open the workspace and explore — no credit card required.

Open vMira