"100k before lunch": a post says the usable context window is a fraction of the box number

The claim: the number on the box is marketing

The post argues advertised context windows (200k, 1M, 2M) are mostly a marketing number, because only a fraction is genuinely usable. There's a "smart zone" of roughly 100k tokens and a "dumb zone" after it where the model quietly degrades. And in real agent work you hit the wall fast: "A few file reads, a long debug session, a sprawling test run, and you're at 100k before lunch."

The author isn't running on vibes alone. He points to two studies: RULER (arxiv 2404.06654), which shows performance dropping across the window, and Chroma's context-rot report, which documents the degradation pattern. The prescription: treat context like a budget, push state out of the live session into written artifacts (PRDs, plans, breadcrumb handoffs) rather than trusting automated summaries.

The practitioners can't agree where the wall is

The thread turned into a scatter plot of personal thresholds, and they're far apart. A core maintainer reported essentially no problem near the top of the window:

"I routinely push past 500k tokens, even sometimes up to around 800k tokens, and don't see this problem … what I see isn't as severe as the author seems to see" — a core Xfce developer, ex-Twilio

The opposite report came just as confidently, the dumb zone arriving far earlier than the author's 100k:

"I find opus models making basic recall mistakes at <100k tokens. Personally I consider < 60k to be the smart zone for opus" — Bolwin

Several converged on the same fix the post recommends without naming it: one conversation per task. A ZeroTier founder and others described pushing all tool-calling into recursive sub-agents so the root thread "can keep the same high level conversation going for an entire day over a million LOC+ codebase" while burning 50M tokens in child calls and never touching 100k at the top.

The objection that the whole thread is anecdote

The sharpest meta-comment pointed at the structure of the discussion itself.

"Almost every comment here is appealing to personal experience. By contrast, OP refers to two studies that compare performance on some kind of standardised test … they can't be worse than anecdotal evidence for something as vague/subjective as LLM performance" — lordgrenville

The reply to that was the load-bearing one: the cited studies are from 2024 and 2025 and may not describe today's models. A commenter who's run RULER-style tests added that Sonnet 4 was "terrible" at the same prompts Sonnet 4.5 nailed, so a benchmark on an old model is its own kind of stale anecdote. And one voice argued context rot has never actually been isolated:

"its existence and severity is purely anecdotal … Context rot is usually taken to mean that the model gets dumber even if it doesn't need to remember specific things … This would be really easy to measure. Just take some standard benchmarks, but fill up the context beforehand" — a commenter proposing the obvious experiment nobody's published

Why the disagreement is real, not noise

The thread looks like chaos but resolves to one point that nobody can wave away: "context window size isn't quite the issue," as one harness-builder put it — the issue is whether the window fills with debris. A long context of clean, relevant material can stay sharp past 800k; the same window polluted with failed attempts and conflicting instructions rots well under 100k. That's why the Xfce dev and Bolwin can both be right. The author's 100k rule isn't wrong so much as it's a number standing in for a quality problem — and as one reply noted, different models use different attention and different compaction, so there's no reason your wall is anyone else's.