What Is AI Context? — The "Reads but Doesn't Read" Reality of the 1M-Token Era
In 2026, Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and DeepSeek V4-Pro all declared "1 million (1M) tokens" of context window. But independent benchmarks (multi-needle NIAH) show that only Gemini 3 Deep Think holds accuracy across the full 1M; the others start losing precision at 200K–400K. "Supports" and "actually reads to the end" are different things. This article walks through how context windows work, the May 2026 model lineup, what Lost in the Middle and Context Rot really are, the cost trap of OpenAI's long-context surcharge, and five practical saving tactics — "cut the session," "send excerpts," "restate at the end," "cache," "explicit addresses" — backed by real benchmark numbers.