Don't trust large context windows
Points and comments are a snapshot, not live.
LLM context windows degrade sharply around 100k tokens regardless of advertised size; treat available context as a budget.
Modern LLMs split into a "smart zone" where performance is reliable and a "dumb zone" where attention degrades, with the boundary around 100k tokens despite vendors advertising 200k to 2M windows. Studies like RULER and Chroma's context rot research show effective context is a fraction of advertised capacity. Coding agents burn tokens quickly and enter degradation before tasks complete. Solutions include auto-compaction (Claude Code) and manually starting fresh sessions with written specifications or structured artifacts like PRDs, plans, and skills. Tools like obra/superpowers structure workflows around small named artifacts to keep live sessions in the smart zone by offloading information to external documents.
What commenters are saying
Thread splits sharply on whether the problem exists. Top commenters report using 500k-800k tokens on Opus and Fable without degradation, contradicting the 100k threshold. Others report degradation as early as 60k tokens and attribute it to "tainted" context from conflicting instructions or bad paths. Several users describe workflow strategies: breaking tasks into overlapping chunks, PRD-based hand-offs between sessions, and design documents with phase-by-phase planning. Disagreement centers on whether context rot is model-dependent (newer Fable performs better than Opus 4.6) or user-dependent (misconfiguration, bad prompts, or scope issues). One user notes that 100k "by lunch" seems inconsistent with modern model performance.