Why I have my orchestrator compact at 100K tokens A developer who runs Claude Opus 5.5 agents with a 1M-token context window reports compacting the orchestrator at 100K tokens, citing observed drift past roughly 200K tokens and warning that orchestrator drift propagates through subagent briefs. The writeup points to Anthropic's session-management guidance and Chroma's July 2025 context-rot evaluation of 18 LLMs, which found performance grows increasingly unreliable as input length grows, and notes that prompt caching cuts the price of 100K input tokens to as little as $0.02 on a cache hit without reducing the token count. It also flags compaction's downsides, including lossy summaries and broken prefix-cache reuse, and proposes a repeatable comparison logging input and cache token counts. Claude Opus 5.5 has a 1M token context window https://platform.claude.com/docs/en/models/opus-5-5/overview . I keep every agent under 100K anyway. Past 200K a slight drift sets in. These are personal observations, not measured thresholds. When a subagent drifts, you can get a weaker answer back. When the orchestrator drifts, that drift can spread through its briefs. Claude Code's subagents normally start with a fresh, isolated context window https://code.claude.com/docs/en/sub-agents . Forks inherit the parent conversation instead. A fresh subagent gets its task through the orchestrator's brief. Without the history, it may miss a constraint the brief dropped. Anthropic's session guide https://claude.com/resources/articles/using-claude-code-session-management-and-1m-context warns that growing context can hurt performance. Claude Code summarizes automatically near the window's limit. The guide recommends compacting earlier and saying what comes next. /context /compact focus on the auth refactor, drop the test debugging /context https://code.claude.com/docs/en/commands shows what's filling the window. The /compact line is the guide's own example. Codex has /compact https://learn.chatgpt.com/docs/developer-commands?surface=cli too. Chroma's Kelly Hong, Anton Troynikov and Jeff Huber https://www.trychroma.com/research/context-rot evaluated 18 LLMs in July 2025. Performance "grows increasingly unreliable as input length grows". The models included GPT-4.1 and Claude 4, not Opus 5.5. Neither source draws a line at 100K or 200K. For 100K input tokens on Opus 5.5, using API prices checked on November 3, 2026 https://platform.claude.com/docs/en/build-with-claude/prompt-caching : | Billed as | Per million tokens | 100K tokens | |---|---|---| | Uncached input | $4 | $0.40 | | Cache write, 5 minutes | $5 | $0.50 | | Cache write, 1 hour | $8 | $0.80 | | Cache hit | $0.20 | $0.02 | Each amount is the listed rate multiplied by one tenth. These are input costs only; output and tool charges are excluded. They are not subscription prices. Other models have other multipliers. The hit is 0.05× the input price, and it doesn't buy room. Anthropic's words https://platform.claude.com/docs/en/build-with-claude/context-windows : "prompt caching changes what you pay for those tokens, not whether they count." Compaction also rewrites the history. Hits need 100 percent identical prompt segments https://platform.claude.com/docs/en/build-with-claude/prompt-caching , so the unchanged start can still hit, and the rewritten part can't. CompactionBench https://www.microsoft.com/en-us/research/publication/benchmarking-long-horizon-agent-context-compaction-with-compactionbench/ , from Microsoft researchers, lists "disrupting prefix-cache reuse" among the problems. Amp removed compaction https://ampcode.com/news/handoff on October 23, 2025 and replaced it with Handoff. It calls compaction lossy: the agent decides what survives in the summary. Compacting early still carries that risk. A constraint you stated early can disappear. This comparison has not been run for this article. /context /compact . Restore the repo snapshot before each run. Repeat each comparison. task,model,effort,input tokens,compacted,run,tests passed,constraints kept On the API, log input tokens , cache read input tokens and cache creation input tokens https://platform.claude.com/docs/en/build-with-claude/prompt-caching . Those separate uncached input from cache reads and writes. The maintained article, with all ten context terms and what each does to the budget, is on argbe.tech https://argbe.tech/blog/agent-context-compaction-100k?utm source=devto&utm medium=syndication&utm campaign=agent-context-compaction-100k .