Context rot is now quantified across 18 frontier models — and coherent input degrades attention worse than shuffled input

Context Rot, RAG, and Long Context: How to Architect LLM Systems in 2026·high signal

Cross-model testing over GPT-4.1, the Claude 4 family, Gemini 2.5, and Qwen3 shows accuracy dropping non-uniformly by 30–50% well before the documented context limit, with semantic-similarity decay mattering more than raw length — counterintuitively, coherent well-structured input degrades attention more than shuffled input. A concrete agent-safety instance: Opus 4.6 acting as a monitor over long agent transcripts on MonitorBench drops from 98.6% to 88% recall as context grows, meaning your oversight layer silently gets worse exactly when the run gets long enough to need it. The emerging 2026 default is hybrid — retrieve 50K–200K relevant tokens and long-context-reason over those, rather than dumping everything in; and if you monitor your own agents with an LLM, chunk the transcript rather than passing it whole.