Member-only story
Inside “context rot” — the 2026 discovery that’s forcing every serious AI team to rethink how they feed information to their models #
For the last three years, the AI industry ran on one assumption: more context is always better. If a model could only read 4,000 tokens in 2023, and 200,000 by 2025, and now 1–2 million tokens in 2026, then surely the fix for every messy AI problem was simple — just stuff more into the window. More documents. More chat history. More tool outputs. Let the model sort it out.
The Discovery: Models Get Worse, Not Better, as You Feed Them More #
In 2025, the AI infrastructure company Chroma ran a study that quietly reshaped how engineers think about large language models. They tested 18 frontier models — including GPT-4.1, the Claude 4 family, Gemini 2.5, and Qwen3 — on simple retrieval tasks, and increased the input length step by step.
Every single model got worse as the input grew. Not some of them. All eighteen.
That result mattered because it broke the mental model most people had been using. A context window isn’t a hard drive with a capacity limit where performance is flat…