How to cut Claude Code costs: --effort, model choice, /usage

The quiet-but-effective cost levers, verified against real CLI help output.

Review usage before upgrading your plan

Before jumping to a higher plan, check how efficiently you're using the current one. Plan comparisons live in the pricing guide; this article focuses on same-plan adjustments.

Tune reasoning depth with --effort

Verified via `claude --help` on this Mac: `--effort <level>` accepts `low`, `medium`, `high`, `xhigh`, `max`. Use low for routine work, high+ for hard design/debugging.

$ claude --help | grep -A2 -- '--effort'
--effort <level>   Effort level for the current session (low, medium, high, xhigh, max)

Match the model to the task with --model

Use `--model` aliases (e.g. `sonnet`, `opus`) to route hard decisions to a stronger model and routine work to a lighter one. `--fallback-model` avoids stalls when your chosen model is unavailable.

Check real usage in /usage

The old `/cost` and `/stats` now open as tabs inside `/usage` — check it regularly to see what's driving spend.

Context management matters too

Long single sessions accumulate context, which costs more per request. Split sessions at natural breakpoints, keep CLAUDE.md lean, and don't leave unused MCP servers/plugins enabled.

Watch for double billing

Your Claude Code subscription and direct Anthropic API usage (from self-built MCP servers/apps) are billed separately — check both.

FAQ

What values does --effort accept?

low, medium, high, xhigh, max (verified via claude --help).

Does model choice affect cost?

Yes — route hard tasks to stronger models, routine work to lighter ones.

Where do I check usage?

`/usage` in a session; the old /cost and /stats are now tabs there.

How do I lighten context?

Split sessions, keep CLAUDE.md lean, disable unused MCP servers/plugins.

Are subscription and API usage billed together?

No, they're separate — check both if you also call the API directly.

See ConfigDeck pricing →Browse the market →See School →