--- title: 1M context windows are a flex. 150K is where the work... description: Large context windows in LLMs often degrade output quality rather than improve it — a phenomenon called 'context rot.' Transformer self-attention scales... canonical: https://daily.dev/posts/1m-context-windows-are-a-flex-150k-is-where-the-work-actually-happens--ydlbsut9j twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: 1M context windows are a flex. 150K is where the work actually happens. | daily.dev og:description: Large context windows in LLMs often degrade output quality rather than improve it — a phenomenon called 'context rot.' Transformer self-attention scales... og:url: https://daily.dev/posts/1m-context-windows-are-a-flex-150k-is-where-the-work-actually-happens--ydlbsut9j og:image: https://api.daily.dev/og/posts/yDlbSut9j.png og:image:alt: 1M context windows are a flex. 150K is where the work actually happens. og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # 1M context windows are a flex. 150K is where the work actually happens. **[Trends](https://daily.dev/sources/trends)** · 3 min read · 2 upvotes · 0 comments ## Summary Large context windows in LLMs often degrade output quality rather than improve it — a phenomenon called 'context rot.' Transformer self-attention scales quadratically with context length, increasing noise and causing the 'lost in the middle' problem where relevant information gets ignored. Sycophantic responses ('You're completely right!') signal a trajectory-poisoned session. A real-world case study shows a team scrapping an entire system after letting an LLM write unreviewed code for four months, resulting in an unauditable codebase. Practical mitigations include aggressive trimming, recursive summarization, reranking context chunks before injection, and starting fresh sessions when drift is detected. The emerging discipline of 'context engineering' — being deliberate about what enters the context, not just how much — is positioned as a key differentiator for developers working with AI systems. ## Content OpenAI just quietly cut GPT-5.6's default context window in Codex CLI by 27% — from 372K to 272K tokens — and the developer community is treating it less as a product decision and more as an accidental confession. The confession: big context windows don't work the way the marketing implies. The underlying problem has a few names right now. "Context rot" describes how output quality degrades as the window fills up with noise — stale messages, redundant tool outputs, contradictory instructions. "The dumb zone" is the more visceral framing: there's a threshold (rough rule of thumb: around 150K tokens) past which models hallucinate more, miss instructions, and generate worse code. The attention mechanism is quadratic, so every token you add costs exponentially more to process, and the model starts losing track of things buried in the middle of the conversation. Gergely Orosz put it plainly: the more context depth you use over longer runs, the more errors compound and the more agents drift. Less context, shorter runs, more reliability. That's the tradeoff nobody advertising 1M token windows wants to lead with. The "lost in the middle" effect is real and well-documented — information placed in the middle of long prompts gets recalled less reliably than information at the edges. Which means that massive context window isn't a flat playing field. It's a cliff with a soft landing at the start and a hard one somewhere around token 150K. Dex Horthy's framing from a recent interview is worth sitting with: "token harder vs. token smarter." The instinct is to throw more context at a problem. The better move is active context management — summarizing history recursively, reranking what actually matters before it goes in, breaking tasks into bounded units with fresh windows. The multi-agent angle is where this gets interesting. The emerging consensus is that more agents aren't automatically better, but each agent with a *smaller, cleaner* context is. One agent holding an entire session is the anti-pattern. Several agents with narrow, focused contexts is the architecture that actually scales. One data point that should make anyone building on AI coding tools nervous: Dex's team shipped unreviewed AI-generated code for months, production broke, and the model couldn't find the root cause. It took three weeks to re-onboard humans to a codebase no human had ever read. That's not a context window problem. But it's what happens when you trust the window more than you should. ## Similar posts on daily.dev - [Context window in AI: why every token is a budget decision](https://daily.dev/posts/context-window-in-ai-why-every-token-is-a-budget-decision-reskx1u2b) · Redis · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#prompt-engineering](https://daily.dev/tags/prompt-engineering) [View this post on daily.dev](https://daily.dev/posts/1m-context-windows-are-a-flex-150k-is-where-the-work-actually-happens--ydlbsut9j) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"1M context windows are a flex. 150K is where the work actually happens.","url":"https://daily.dev/posts/1m-context-windows-are-a-flex-150k-is-where-the-work-actually-happens--ydlbsut9j","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/1m-context-windows-are-a-flex-150k-is-where-the-work-actually-happens--ydlbsut9j"},"datePublished":"2026-07-20T10:25:32.287Z","dateModified":"2026-07-24T22:13:49.550Z","description":"Large context windows in LLMs often degrade output quality rather than improve it — a phenomenon called 'context rot.' Transformer self-attention scales...","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/1m-context-windows-are-a-flex-150k-is-where-the-work-actually-happens--ydlbsut9j","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,prompt-engineering","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"1M context windows are a flex. 150K is where the work actually happens."}]} ```