What a million tokens of context is actually for

A million tokens is roughly 3,000–4,000 pages of text. That's not a bigger chat box — it's a different way of working.

Where it changes the game

1. Whole-repository reasoning. Paste an entire codebase and ask for an architecture review or a cross-file refactor plan. No embedding index, no retrieval.
2. Long-document synthesis. Contracts, research corpora, a quarter of support tickets — analysed in one pass, so cross-references actually resolve.
3. Agent memory. A long-horizon agent can keep its full trajectory in-context instead of summarising and forgetting.

Where it's the wrong tool

• Latency-sensitive single questions — you pay for the context you send.
• Frequently-changing knowledge — a huge static prompt goes stale; retrieval is better when freshness matters.

A rule of thumb

Use the big window when the relationships between distant parts of the input are the point. If you just need a fact lookup, you don't need a million tokens.