One million tokens sounds like a marketing number. Here's what it actually enables.
The Math
One million tokens is roughly:
- 750,000 words
- A 2,500-page book
- About 50,000 lines of code
- 3-4 months of daily email at average volume
For most individual tasks, you don't need anywhere near 1M tokens. But there's a category of tasks where context size was the binding constraint — and that constraint is now effectively gone.
What Becomes Possible
Full codebase analysis: Load an entire medium-sized codebase (50-200K lines) in a single context. Ask architectural questions, trace dependencies, understand the full system. Previously this required chunking, which meant losing cross-file connections.
Long document work: Legal contracts, technical specifications, research papers, book manuscripts — entire documents with all their nuance and interconnection, processed in one context rather than piecemeal.
Extended conversation history: Months of previous conversation and project context loaded into a single session. Your agent genuinely knows the full history.
Multi-document synthesis: Combine multiple reports, papers, or datasets that together exceed previous context limits.
The Practical Tradeoff
Larger contexts are slower and more expensive than smaller ones. You wouldn't use 1M tokens for a quick question. But for the class of tasks where context size was the blocker, the option now exists.
The Competitive Significance
Gemini 2.0 Pro has 2M token context. GPT-5.4 has 1M in specific modes. The context race is largely over — all frontier models are now at a scale where most practical tasks fit comfortably. The differentiation is shifting to what you can actually do with large context, not whether you have it.