LLM context window comparison

How big is your context, really?

A token is roughly three-quarters of an English word —context_words ≈ context_tokens × 0.75. So a 200K token window holds about 150,000 words: one full-length novel, or every README in a mid-sized monorepo, or roughly 600 pages of dense technical PDFs. A 2M token window starts to feel ridiculous — that’s the full Harry Potter series with half of it left over.

Context size matters in three places. RAG depth: the more chunks you can stuff in, the less your retrieval has to be perfect. Agent step count: every tool call, observation, and reasoning step accumulates — agents that think for 20+ turns burn through 32K windows before they finish the task. Long-document work: summarizing a contract, refactoring across files, or answering questions over a research corpus is bottlenecked by the longest single thing you can show the model.

Bigger isn’t free. Most models charge per input token, so a 1M-token prompt at $3/M costs $3 per call before you’ve read a word of output. And recall degrades past ~64K tokens for most models — “needle in a haystack” benchmarks hide the fact that real reasoning over 500K tokens is still hit-or-miss. Pick the smallest window that fits your job.