Context rot hits 18 frontier models before limits

Accuracy drops 30 to 50% across 18 frontier models before any hit their documented context limits, per new testing on GPT-4.1, Claude, Gemini, and Qwen3.

Opus 4.6, used as a monitor watching long agent transcripts, drops from 98.6% to 88% recall as context grows, on the MonitorBench benchmark. That's the oversight layer getting worse exactly when a long-running agent most needs watching.

The degradation isn't uniform, and it isn't just about length. Semantic-similarity decay predicts the drop better than raw token count does. Counterintuitively, coherent, well-organized input degrades attention worse than shuffled input, per the testing.

If you monitor your own agents with an LLM, chunk the transcript before handing it over. Don't pass the whole thing and trust a clean read. That's the failure mode to check for before you trust a monitor's clean report.

What to do

Retrieve 50K to 200K relevant tokens and reason over that narrower set, instead of dumping everything into the window. That's the pattern forming for 2026.