How do I check token usage in Claude Code?

Run /usage for session tokens, estimated cost and plan usage bars, and /context to see what fills the context window. A custom status line can show both on every prompt.

Why does Claude Code use so many tokens?

It re-sends the whole conversation with every request and every tool call. In our logs 96% of tokens were cache reads of existing context, and output was under 1%.

What is the fastest way to reduce Claude Code token usage?

Use /clear between unrelated tasks and lower the effort level. On Opus 5.5 a task cost $5.98 at max effort and $1.34 at medium.

Does /compact reduce Claude Code costs?

It shrinks later requests, but compacting reads the whole conversation, so it is a large request itself. If you do not need the history, /clear costs nothing.

Quick answer: Run /usage to see what Claude Code is spending, and /context to see what is filling the window. To spend less, keep context small: clear between tasks, lower the effort level, and keep verbose output...

How to check and reduce Claude Code token usage

Q: How much does Claude Code cost per developer?

Anthropic reports about $13 per developer per active day and $150 to $250 a month across enterprise deployments at API rates, with 90% of users under $30 a day.

Updated September 27, 2026 · first published September 27, 2026

Run /usage to see what Claude Code is spending, and /context to see what is filling the window. To spend less, keep context small: clear between tasks, lower the effort level, and keep verbose output out of the main conversation. Context matters more than anything else because Claude Code re-sends the whole conversation with every request. In our own logs, 96% of tokens were the agent re-reading its context.

This page covers the commands that show your usage, the settings that move it most, and where the tokens actually go. The tips come from Anthropic’s cost documentation. The numbers come from our own Claude Code logs and published benchmarks.

How do I check Claude Code token usage?

/usage: shows token counts and an estimated cost for the current session, a prompt cache line, and on Pro, Max, Team and Enterprise plans your plan usage bars. It also breaks down what is using your plan: skills, subagents, plugins, MCP servers and scheduled loops, over the last 24 hours or 7 days.
/context: shows what is taking up the context window right now, with hints on what to trim.
/insights: writes an HTML report on your recent sessions and where you lose time and tokens.
The status line: a script can show cost, context used and your 5-hour and 7-day plan percentages on every prompt. See status line docs.
claude.ai Settings > Usage: plan limits and usage-credit spend, including use from other devices.

On a Pro or Max plan, the dollar figure in /usage is what the session would have cost at API list prices, not what you pay. It is still the best single measure of how heavy a session is.

Where do Claude Code tokens actually go?

Mostly into re-reading context. Every request carries the full conversation, and every tool call sends another request with the results. Across our Claude Code logs, 7.0 billion cache-read tokens were 96% of all tokens and half of a $6,986 API-equivalent bill. Output tokens, the part you see, were under 1% of volume.

That is why a one-line question in a session that has been open all day is not cheap. It carries the whole day with it. Our tokens per plan page has the full breakdown: about 1.9 billion tokens in a Max 20x week and about 39 million in a Max 5x session.

How do I reduce Claude Code token usage?

These are the changes that move usage most, in rough order of effect.

Clear between unrelated tasks. /clear starts a fresh context and costs nothing. Use /rename first if you want to /resume the old session later.
Lower the effort level. On Opus 5.5, one task cost $5.98 at max effort and $1.34 at medium in Artificial Analysis runs, a 4.5x difference from one setting. Use /effort to change it. Details on our effort level cost page.
Match the model to the job. Anthropic recommends Sonnet for most coding and Opus for hard architectural or multi-step work. Switch with /model, and set model: haiku on simple subagents.
Don’t take long breaks mid-session. The prompt cache lasts an hour on a subscription and five minutes on usage credits or an API key. Coming back after it expires re-processes your whole context at full price.
Keep noisy output out of the main context. Send test runs, log reading and doc fetching to subagents so only a summary comes back. A hook can filter a 10,000-line log down to the error lines before Claude sees it.
Trim what loads at start. Keep CLAUDE.md under about 200 lines and move workflow-specific instructions into skills, which load only when used. Disable MCP servers you aren’t using with /mcp. Prefer CLI tools like gh where one exists.
Write specific prompts and plan first. “Improve this codebase” makes Claude scan everything. Plan mode (Shift+Tab) catches a wrong approach before it costs a full implementation.
Watch background work. Scheduled loops, agent teammates and idle subagents keep sending your full context. Agent teams use about 7x the tokens of a normal session when teammates run in plan mode.

Does /compact save tokens?

Only on later requests, and compacting a big context is itself a big request. /compact reads the whole conversation to summarize it. If you don’t need the history, /clear is cheaper. If you do, /compact with instructions (for example /compact keep the API changes and failing tests) keeps what matters. You can also set an earlier auto-compact threshold with /autocompact.

How much does Claude Code cost per developer?

Anthropic says about $13 per developer per active day and $150 to $250 a month across enterprise deployments, and under $30 a day for 90% of users. That is billed at API rates. On a subscription the question becomes which plan’s limits you hit first. Our Claude Max weekly limit page prices a Max 20x week at about $1.86k of API-equivalent usage.

How do teams track Claude Code usage?

Team and Enterprise plans: the spend report in org analytics, with a per-user CSV. Enterprise also has an analytics API. Spend limits are set in admin settings.
API (Claude Console): the Console usage page and workspace spend limits.
Any setup: set CLAUDE_CODE_ENABLE_TELEMETRY=1 to export per-user token and cost metrics over OpenTelemetry to your own stack. See monitoring docs.

If you pay contracted rates, the modelPricing managed setting makes the cost figures in /usage and OpenTelemetry match your contract instead of list price.

Sources

Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research