LLM Context Window Usage Calculator (Tokens %)

A context window is the hard ceiling on how many tokens a model can attend to in a single request, and it is shared among the system prompt, conversation history, retrieved documents, the user message, and the generated reply. Tokens are subword units produced by a byte-pair encoding tokenizer, so the character-per-token ratio varies with content: ordinary English runs near 4 characters or 0.75 words per token, while source code, JSON, and text in Chinese, Japanese, or Korean consume noticeably more tokens per character. The usage figure here is estimated prompt tokens plus system prompt tokens divided by the window size, with reserved output subtracted afterwards to show the headroom that is genuinely still available. Two distinctions cause most planning errors. Window size and maximum output tokens are separate limits: a model may accept 128,000 tokens of input yet cap generation at 4,096 or 8,192, so free window space does not convert into a longer reply. And multi-turn cost is cumulative, because resending the full transcript on each turn means the request grows by roughly the sum of every prior message and response. Character-based estimation is approximate and can be off by 10 to 20 percent on technical text, so use the model's own tokenizer, such as tiktoken for OpenAI models, whenever a request sits close to the limit. Retrieval of relevant chunks, summarisation of older turns, and sliding windows are the standard remedies once usage passes roughly 70 percent.

Both halves of this calculation have datable origins. Byte-pair encoding was published by Philip Gage in 1994 as a simple data compression algorithm, and Rico Sennrich, Barry Haddow, and Alexandra Birch at the University of Edinburgh repurposed it for neural machine translation in 2015 to handle rare and unseen words by splitting them into subword pieces. That adaptation became the standard tokenization scheme for transformer language models, which is why usage and billing are denominated in tokens rather than words or characters. Window sizes grew as the cost of attention was engineered down. The original transformer described by Ashish Vaswani and colleagues at Google in 2017 trained on sequences of a few hundred tokens, because self-attention cost grows with the square of sequence length. GPT-2 in 2019 used 1,024 tokens and GPT-3 in 2020 used 2,048. Efficiency work then compounded: rotary position embeddings introduced by Jianlin Su and colleagues in 2021 gave positional encoding better extrapolation, and FlashAttention, published by Tri Dao and co-authors at Stanford in 2022, made exact attention far more memory-efficient. By late 2023 commercial models offered 128,000-token windows, Anthropic shipped 200,000 tokens for Claude, and Google announced a one-million-token window for Gemini in 2024, moving the practical constraint away from raw capacity toward cost and retrieval quality.