Javlon Baxtiyorov
← Writing

A 1.5-million-token context window is a design problem, not a feature

OpenAI's GPT-5.6 reportedly pushes context to 1.5M tokens and trims another 10–15% off the token bill. Bigger context doesn't free you from retrieval discipline — it moves where the discipline lives.

A 1.5-million-token context window is a design problem, not a feature
Photo by Zoshua Colah on Unsplash

OpenAI's GPT-5.6, expected late June 2026, reportedly extends the context window to 1.5M tokens and trims token usage another 10–15% over GPT-5.5, with the emphasis squarely on agentic workflows. The headline writes itself: "you can just put everything in the prompt now."

You can't. Or rather — you can, and you'll regret the bill, the latency, and the quietly degraded answers.

What a bigger window actually changes

A larger window doesn't repeal the rules; it moves them:

  • Relevance still beats volume. A model handed 1.5M tokens of mostly-irrelevant context answers worse, not better, than one handed the right 8K. "Lost in the middle" doesn't go away because the middle got longer.
  • The cost is real even when it's cheaper. A 10–15% per-token saving against a window you've made 10× larger is not a saving. Efficiency gains get eaten by capacity the moment you stop rationing.
  • Retrieval becomes curation. The job shifts from "fit it in the window" to "decide what deserves to be there." That's still retrieval. It's still my problem.

How I build around it

The bigger window earns its keep in specific places — a long agent trajectory that needs its own history, a whole file under review, a multi-document synthesis where chunking was lossy. For those, it's genuinely better. For everything else, I still retrieve narrowly and pass the minimum.

The trap with every capability jump is treating it as permission to stop designing. A 1.5M-token window is a sharper tool, not a license to skip the thinking. I'll reach for it where the task is genuinely long-horizon, and keep my context lean everywhere else — because the model that reads everything also pays for everything, and so do I.


Sources: Introducing GPT-5.5 (OpenAI), OpenAI plans June GPT-5.6 (AI Weekly).


← All writing Get in touch →