Compacting at 500k tokens costs a third less than 1M: real usage data on context compression
EricBuess · x · 2026-11-14
Eric Buess shares real cost data on context compaction: setting compaction around 500k tokens costs roughly a third less per day than at 1M, with diminishing savings below that. His chart shows daily compactions drop from 27 (250k) to 4 (1M) while plan points rise from 19 to 31; each compaction reloads 110-130k tokens and rewrites 80-100k. Cached re-reads cost 1/20 of a fresh token yet still account for 60% of plan points. He also notes models follow the thread worse past half the window, and uses Opus 5.5 headless voice sessions with hot swapping instead of compaction.
More from coding & agent
- Yogurt: open-source local-first Granola alternative for private AI meeting notes — supportingthedogs · 2026-11-14
- Anthropic Breaks Down the Five Primitives Behind Always-On AI Agents — dotey · 2026-11-14
- Context Vim idea: diffusion LLMs could edit full code pages with control tokens — ryunuck · 2026-11-14
- Agent drives Unreal Engine via MCP: auto lighting, rendering and self-critique loop — HankYeomans · 2026-11-14
- Running Claude agent sessions visualized as characters in an Oasis game world — LeviTurk · 2026-11-14
- Vibe-Coded Game Hits 35.7K Plays and 10.5K Players in First Week on Spawn — TAbrodi · 2026-11-14