Restart Claude Code Sessions at 200K Tokens (Save Up to 80%) (2026)

The fix: when your session crosses 200K tokens (ccusage shows this), stop, ask Claude to write a handoff doc (what is done, what is next, relevant file paths), run /new, paste the handoff, and continue. One developer documented an 80% monthly drop — $1,600 to $320 — from this single behavior change. Restart discipline is the highest-ROI habit on the whole cost list.

Why long sessions get expensive

Claude Code is stateless per call: each message resends the entire conversation so far as context. Early in a session that is cheap. But context grows monotonically, and past 200K tokens the per-message cost climbs sharply — a 300K+ session can cost 10-15x more than the same work split across fresh sessions. You are paying to re-read the whole history on every turn.

The restart pattern

1. When ccusage shows your session crossing 200K, stop.
2. Ask Claude to write a handoff doc: what is done, what is next, relevant file paths, and any open questions.
3. Save it, run /new, paste the handoff, and continue where you left off.

The handoff takes about 30 seconds to produce and reloads only the context that matters, instead of dragging 300K tokens of history into every future message.

Automate the ceiling

You can also force summarization earlier in ~/.claude/settings.json:

{
  "compactThreshold": 200000,
  "autoCompact": true,
  "subagentModel": "claude-haiku-4-5"
}

compactThreshold: 200000 compacts context at 200K instead of the default 300K+, and subagentModel routes spawned sub-tasks to Haiku. Reports put this at a further 30-50% on top of the manual restart discipline.

OpenClaw users: the same settings.json applies to agents routed through the gateway. Keep OpenClaw sessions from ballooning by using /new between unrelated tasks and letting autoCompact cap the context.

Related cost guides

• OpenClaw Costs Guide (all 10 fixes) — the full cost-optimization hub
• ENABLE_TOOL_SEARCH: cut the 45K-token tool tax — trim per-turn context
• Caveman Mode for Claude Code — cut output tokens 61-75%

Need help?

If your bill is bleeding and you would rather not tune this yourself, we apply the whole checklist and set up ccusage monitoring in one session. See training and cost-audit options →