Last reviewed:
What is a context window? Definition and business implications
The context window is the maximum amount of text an AI model can process in a single query, measured in tokens. It includes the prompt sent and the response to produce. Beyond this limit, the model loses access to the oldest information of the exchange.
The context window is a technical characteristic of the model, fixed at construction. For GPT-3 in 2020, it was 2,048 tokens, that is, about 1,500 words or 3 pages. In 2026, the main providers offer one million tokens or more (between 550,000 and 750,000 words depending on the tokenizer, or over 1,000 pages): this is the case for all four current Anthropic Claude models, Haiku 5.5 included, OpenAI's GPT-6 models and DeepSeek's V4 models. Meta even announces ten million for Llama 4 Scout. This expansion radically changes the possible use cases. You can inject a complete dossier, an entire knowledge base, a conversation history of several months. But beware: declared size and effective quality diverge. Research on long contexts shows it: precision drops when the relevant information sits in the middle of a very long context (Liu et al., Stanford, 2023), and more broadly as the context grows (Chroma, 2025). For critical cases, a tight, well-structured context is preferable to an 800,000-token dossier dumped in bulk.
Concrete example
In December 2022, GPT-3.5 accepted 4,096 tokens as input, that is, about 3,000 words. In March 2026, Anthropic released Claude Sonnet 4.6 in general availability with a one-million-token window, at standard pricing (3 dollars per million input tokens, 15 per million output). In three years, single-call ingestion capacity has been multiplied by 244. Concretely, a law firm can submit the entirety of a client file (contracts, exchanges, case law) in a single prompt, where it would have previously needed to orchestrate a RAG with vector database. The architectural simplification is radical, the cost also kept in check.
FAQ
What is the context window?
It is the maximum amount of text an AI model can process in a single query, measured in tokens. It includes both the prompt sent and the response to produce. Beyond this limit, the model loses access to the oldest information of the exchange.
How large is it in 2026?
One million tokens or more at the main providers, between 550,000 and 750,000 words depending on the tokenizer, or over 1,000 pages, whereas GPT-3 was limited to 2,048 tokens in 2020.
Does a large context window replace a RAG?
Often, yes. Use cases that yesterday required a complex RAG architecture (large client file, whole documentary base) now fit in a single call to a standard LLM. Before funding a RAG, check whether a long context suffices.
Does a large window guarantee answer quality?
No. Declared size and effective quality diverge: precision drops when the relevant information sits in the middle of a very long context (Liu et al., 2023), and more broadly as the context grows (Chroma, 2025). On the NoLiMa benchmark, 11 of 13 models fall below half of their baseline score at 32,000 tokens. For critical cases, place the essentials at the start and end of the prompt, not in the middle.
Is an abundant context free?
No. Inference cost stays proportional to token volume: a 500,000-token prompt costs 500 times more than a 1,000-token one. With some providers, very long requests are even billed at a higher unit rate.
See also
Sources
- Context windows, Anthropic documentation (sizes by model). https://platform.claude.com/docs/en/build-with-claude/context-windows
- 1M context is now generally available for Opus 4.6 and Sonnet 4.6, Anthropic, 13 March 2026. https://claude.com/blog/1m-context-ga
- OpenAI API, Models: context window of GPT-6 models, September 2026. https://developers.openai.com/api/docs/models
- Language Models are Few-Shot Learners (GPT-3), Brown et al., arXiv:2005.14165, 2020. https://arxiv.org/abs/2005.14165
- The Llama 4 herd, Meta AI, April 2025. https://ai.meta.com/blog/llama-4-multimodal-intelligence/
- What is a context window?, IBM Think 2026. https://www.ibm.com/think/topics/context-window
- Lost in the Middle: How Language Models Use Long Contexts, Liu et al., Stanford, 2023. https://arxiv.org/abs/2307.03172
- Context Rot: How Increasing Input Tokens Impacts LLM Performance, Chroma, 14 July 2025. https://www.trychroma.com/research/context-rot
- RULER: What's the Real Context Size of Your Long-Context Language Models?, Hsieh et al., 2024. https://arxiv.org/abs/2404.06654
- NoLiMa: Long-Context Evaluation Beyond Literal Matching, Modarressi et al., 2025. https://arxiv.org/abs/2502.05167