LLM Context Window Comparison

The context window is the maximum number of tokens a model can process in a single request — system prompt, conversation history, retrieved documents and the model’s own output all count against it. Exceed it and the API either truncates the oldest content or returns an error. Bigger windows let you stuff entire codebases, long documents, or multi-hour agent histories into one request.

Gemini 3.5 Flash currently leads with 1M tokens — roughly 750K words or about 1,500 pages of text. Google’s Gemini lineup ships 1M across the family, OpenAI’s GPT-5 family offers 400K, and Anthropic’s Claude models offer 200K.

About 96,000 words — or roughly 190 pages. English text averages 1.3 tokens per word (0.75 words per token), so a 128K-token window holds ≈ 96K words. Code is denser: expect 30–50% fewer "useful" tokens after comments and whitespace. Our GPT-5 entry holds 300K words at 400K tokens.

The window itself is free — you pay per token actually sent, not for the capacity. But in practice big contexts are expensive: sending ${fmtTokens(largest.contextWindow)} input tokens to fill Gemini’s window costs ${fmtCost((largest.contextWindow / 1_000_000) * (largest.inputPer1M ?? 0))} at list price, every single request. Prompt caching (10% of normal input cost on Claude and GPT-5) is how production systems survive long contexts.

The API rejects the request with a context-length error — it does not silently truncate in most cases. You either trim history, summarize older turns, switch to a bigger-window model, or move to a RAG setup that retrieves only relevant chunks instead of stuffing everything.

Not necessarily. Effective context use often degrades deep into very long inputs — models can miss details buried in the middle of huge prompts ("lost in the middle" behavior). For retrieval-heavy workloads, a well-chunked RAG pipeline over a 200K model frequently beats dumping 1M tokens of raw text. Pick the window that fits your real document sizes, not the biggest number on the box.