Keeping Claude Code usage and costs under control

Claude Code costs stay under control when you do three things: watch where usage goes with /usage, keep each session small and focused, and set hard limits on anything that can bill beyond your plan. Most runaway usage comes from long sessions that are never cleared, the most capable model left on for simple edits, vague prompts that make Claude read half the site, and background work you forgot was running. All of those are habits and settings, not plan upgrades.

This guide is about using less and capping spend. It doesn't repeat plan prices, which our Claude Code pricing guide covers in full. In short, as of October 2026: subscriptions (Pro, Max, Team) include a usage allowance with five-hour and weekly limits, shared with Claude chat; usage credits and API accounts bill per token at API rates. A token is a small chunk of text; models are metered by the tokens they read and write. The controls below apply to both, with notes where they differ.

Where the money actually goes

On every route, cost scales with context: everything Claude is holding in the session, including your messages, the files it has read and command output. Anthropic's costs page: "Token costs scale with context size: the more context Claude processes, the more tokens you use." Every request sends the conversation so far, so a long session makes every new message more expensive.

Read /usage before you change anything

Type /usage in a session (/cost and /stats are aliases). On Pro, Max, Team and Enterprise it shows your plan usage bars and a breakdown of what counts against your limits:

• Attribution: recent usage attributed to skills, subagents, plugins and individual MCP servers, each as a percentage.
• Behavior flags: patterns such as long context or cache misses, flagged when one accounts for 10% or more of recent usage, each with a tip.
• Loops: the heaviest scheduled tasks that ran recently, with tokens per run.

Press d or w to switch between the last 24 hours and the last 7 days. The figures come from session history on that machine, so usage from other devices or claude.ai isn't included. If usage credits are on, a row shows this month's spend against your limit. For API users, the Session block shows an estimated dollar cost; Anthropic points to the Console's Usage page for authoritative billing.

For a pattern-level view, /insights writes a report on how you work, including friction points such as misunderstood requests. Note that the analysis itself uses tokens from your plan or API usage.

Session habits that use less

1. One task per session. Run /clear when you switch to unrelated work. Use /rename first if you want to come back to it with /resume.
2. Write specific prompts. Anthropic: vague requests "trigger broad scanning," while specific ones let Claude "work efficiently with minimal file reads." "Fix the meta description on /pricing/" costs less than "improve our SEO." Our guide to writing prompts for website changes shows how.
3. Plan before big jobs. Plan mode has Claude propose an approach before editing, which Anthropic says prevents "expensive re-work when the initial direction is wrong."
4. Stop early. Press Esc as soon as Claude heads the wrong way, and rewind rather than arguing it back on course. After two failed corrections, clear and start again with a better prompt.
5. Give it a check. A build, a test or a target screenshot lets Claude catch its own mistakes before you have to ask for fixes.
6. Use side questions. /btw answers a quick question without adding it to the conversation history.

When a session does need to keep going, /compact with a focus, such as "focus on the files changed so far," tells Claude what to keep when it summarizes. Our guide to context, compaction and memory explains what survives.

Match the model and effort to the job

As of October 2026, Claude Code defaults to Opus 5.5 on Pro, Max, Team, Enterprise and the Anthropic API. Anthropic's advice is to "reserve Opus for complex architectural decisions or multi-step reasoning" and notes that Sonnet "handles most coding tasks well and costs less." On the API, Sonnet 5.5 lists at $2 per million input tokens and $10 per million output tokens, against $4 and $20 for Opus 5.5.

• Switch models with /model. It saves your choice as the default; press s on a row to switch for the current session only.
• Lower the effort level, which sets how much the model reasons before answering, with /effort for simple, repetitive edits. Thinking tokens bill as output tokens, and you can't turn thinking off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models.
• Give subagents a smaller model. A switch to Opus also applies to subagents that inherit your session's model. For simple subagent jobs, Anthropic suggests model: haiku in the subagent's settings.

Try a cheaper model on one representative task and compare the result before you switch for good. Quality matters more than the per-token price if the cheaper result needs redoing.

Watch the work that runs when you're not looking

Anthropic's costs page lists reasons usage climbs in a long session even when you aren't typing:

Anthropic notes that ordinary background jobs, such as summarizing conversations for resume, typically cost under $0.04 per session. The items above are the ones that grow. Our guides to subagents and routines and scheduled tasks cover when they're worth it.

Keep CLAUDE.md and tools lean

• Short CLAUDE.md. This project instructions file loads into every session. Anthropic suggests keeping it under 200 lines and moving workflow-specific instructions into skills, which load only when used. Our CLAUDE.md templates stay well under that.
• Fewer MCP servers. Tool definitions are deferred by default, but each connected server still adds to context. Run /mcp and disable servers you aren't using.
• Prefer command-line tools where they exist. Anthropic says tools such as gh are "more context-efficient than MCP servers because they don't add any per-tool listing."
• Check what's filling context with /context, which shows usage as a grid with optimization suggestions.

Set hard limits

Habits reduce usage. Limits cap it.

• Pro and Max: if you turn on usage credits, set a monthly spend limit in Settings, Usage on claude.ai. In Claude Code, /usage-credits opens that page. You can also leave credits off and wait for limits to reset; with automatic continue, Claude Code can pick up an interrupted task on its own after the reset.
• Team and Enterprise: the seat allowance is the default ceiling. If admins turn on usage credits, they can set spend limits at the organization, group or member level. Members without billing access can request more from inside Claude Code.
• Claude Console: set workspace spend limits. Claude Code creates its own "Claude Code" workspace the first time you sign in with a Console account, which makes its spend easy to track.
• Cloud providers: use your cloud account's budget controls.

Avoid three surprise bills

1. An API key in your environment. If one is set, Claude Code can use it instead of your subscription, billing per token. Check /status to see which account a session uses.
2. Assuming plan API credits cover Claude Code. The monthly API credits that come with Max and Team don't cover interactive Claude Code in the terminal, IDE, desktop or web.
3. Automatic pull request review. Anthropic's managed Code Review for Team and Enterprise, a research preview, is billed separately through usage credits, averages $15 to $25 per review, and can run after every push, "multiplying cost by the number of pushes." Set a spend cap before turning it on.

A monthly cost review

1. Each person runs /usage, switches to the 7-day view, and notes the top attribution items and any behavior flags.
2. Admins export the spend report (Team and Enterprise) or check the Console usage page (API).
3. Look for the usual causes Anthropic names for high spend: "long sessions that were never cleared" and "Opus left as the default model."
4. Stop loops, routines and MCP servers nobody uses.
5. Upgrade only the people who regularly hit limits, and only after their habits are fixed.

Keep the subscription in proportion. For most website teams, the bigger cost is people's time writing tasks and reviewing changes. Our guide to measuring Claude Code ROI shows how to count both.

Conclusion

Keeping Claude Code costs under control is mostly routine: read /usage weekly, clear between tasks, write specific prompts, match the model and effort to the job, switch off background work you don't need, and put spend limits on anything that bills past a plan. Anthropic's own advice is to start with a small pilot and establish a baseline before rolling out wider. If you'd like help setting up plans, limits and team habits for your website work, our Claude Code consulting covers rollout and training. More guides are in our Claude Code Guide.

Frequently asked questions