LLM Context Windows Explained — Size, Cost, and Trade-offs
A context window is the maximum input + output length a model can process at once. Modern models range from 8K to 2M tokens. Larger windows are powerful but come with cost, latency, and quality trade-offs.
What the context window is
The context window is a hard limit on the total number of tokens a model can "see" in a single request — input tokens (your prompt, documents, conversation history) plus output tokens (the response). If you exceed it, the API returns an error or truncates.
Current context window sizes
Claude Opus 5 / Sonnet 5: 200K. GPT-5.4: 128K. GPT-5 mini: 128K. Gemini 2.5 Pro: 1M (2M extended). Gemini 2.5 Flash: 1M. Llama 4 Scout: 10M (experimental). For most tasks, 128K is more than enough — 128K tokens ≈ 300 pages of text.
The lost-in-the-middle problem
Research shows LLMs perform best on content at the start and end of the context window. Content in the middle of a very long prompt may be "lost" — the model ignores or misremembers it. For document analysis over 50K tokens, test recall accuracy before assuming the model reads every word.
Longer context = higher cost
You pay for every token in your input. A 100K-token document context costs 100× more to process than a 1K-token query. Before stuffing context, ask: can retrieval (RAG) give the model only the relevant 5K tokens instead of the full 100K? RAG is often 10–20× cheaper for large document corpora.
Related calculators
Related guides
Frequently Asked Questions
Which model has the largest context window?
Gemini 2.5 Pro supports 2M tokens (experimental). Llama 4 Scout claims 10M (experimental, performance may degrade before that limit). For reliable large-context processing, Claude's 200K and Gemini's 1M are the proven production options.
Is a larger context window always better?
No. Larger context costs more (linearly), takes longer to process, and quality can degrade on very long inputs due to the lost-in-the-middle problem. Use the minimum context needed for your task. RAG is often superior to long-context for large document corpora.
Guide updated 2026-08-20. Pricing data verified weekly. See methodology · sources.