Cost & Context Optimization

Keeping token spend and context usage low — through model choice, prompt caching, scoping, and clearing — so long sessions stay cheap and sharp instead of bloated and expensive.

Why it matters

Cost and context are the same problem: every token in the window is paid for on every turn, and a stuffed window also degrades reasoning. A careless session re-reads huge files, never clears, and runs Opus for trivial edits; a disciplined one caches the stable prefix, scopes reads, and picks the model per task. The difference is routinely an order of magnitude on the bill — and noticeably better output.

How it works

Pull the levers in roughly this priority order:

Lever	Mechanism	Typical win
prompt caching	reuse stable prefix (CLAUDE.md, tools)	up to ~90% off cached input
right model	Haiku/Sonnet for routine, Opus for hard	large per-token cost gap
scope context	.claudeignore, targeted reads	fewer input tokens per turn
/clear	drop stale history between tasks	resets window to ~0
/compact	summarize history, keep the thread	shrinks without full reset

• Caching keys on an identical leading prefix, so a stable claude-md-project-memory and tool set pays off across turns; volatile content at the front breaks the cache and quietly raises cost.
• Cached reads are billed far below fresh input; the loss leader is the first uncached turn, so longer sessions amortize better — until the window fills.
• Scope what enters context: claudeignore-file-scoping keeps node_modules/build dirs out, and asking for specific files beats “read the whole src tree.”
• /clear between unrelated tasks is the highest-leverage habit — it both cuts cost and removes distractor context; /compact preserves continuity when you can’t fully reset. See managing-context-clear-compact.
• In headless runs, enforce ceilings with --max-budget-usd and --max-turns; pin --model so an upstream default change can’t inflate spend.

Example

Expensive habit:  one Opus session, never cleared, re-reads the 4k-line module every turn.
Cheaper pattern:
  - Sonnet for the implementation, Opus only for the gnarly design step
  - .claudeignore excludes dist/ and vendor/
  - /clear after each feature so history doesn't compound
  - cached CLAUDE.md prefix reused across the whole session

Same task, a fraction of the tokens — and the model isn’t wading through stale context.

Pitfalls

• Cache-busting prefix. Putting timestamps or per-turn data at the start of context invalidates the cache every turn, erasing the discount.
• Never clearing. A day-long session accumulates dead context you keep paying for and that dilutes attention — /clear aggressively.
• Opus for everything. Defaulting to the top model on trivial edits wastes money; match the model to task difficulty (see choosing-a-model).
• Compact at the wrong time. /compact mid-task can summarize away a detail the next step needed; clear between tasks, compact within a long one.

See also

• choosing-a-model
• multi-step-agentic-tasks
• claudeignore-file-scoping