How to read /usage and /context in Claude Code

Claude Code, Anthropic's coding tool, has two commands that work as measuring tools. /usage reports how many tokens the session has used so far. /context reports what is in the context window at this moment. A token is a piece of text that the model processes, and the context window is the largest amount of text the model can read in one request.

The two commands answer different questions, and reading the wrong one leads to the wrong fix. This page explains each line of both screens, using Anthropic's documentation as read on 4 October 2026. It belongs to the guide on how to reduce Claude Code token usage. Our advice is short: measure before you change a setting, and measure again after.

Which screen answers which question

Your question	Where to look	What you get
How many tokens has this session used, and what would they cost?	/usage, the Session block	Token counts for each model and an estimated cost at list price [1]
Is the prompt cache working?	/usage, the Prompt cache (main) line	Share of input read from the cache, and the number of misses [1]
What counts against my plan limit?	/usage, the plan usage breakdown	Shares for skills, subagents, plugins and MCP servers [1]
Which repeating task uses the most?	/usage, the Loops rows	Total tokens and tokens per run for each scheduled task [1]
What is filling the context window now?	/context	A grid of current usage by category, with suggestions [3]
How full is the window, without asking?	The status line	A percentage that updates after each response [2]
How do I work, and where do requests go wrong?	/insights	A report on your recent sessions [1]
What will the organisation be charged?	The Usage page in the Claude Console	The billing record, which Anthropic calls authoritative [1]

The Session block of /usage, line by line

Type /usage in a session. Anthropic's command reference, read on 4 October 2026, describes it as showing "session cost, plan usage limits, and activity stats", and lists /cost and /stats as other names for the same command [3]. The Session block is at the top. This is the sample that Anthropic's documentation prints [1]:

Total cost:            $0.55
Total duration (API):  6m 20s
Total duration (wall): 6h 33m 10s
Total code changes:    0 lines added, 0 lines removed
Usage by model:
   claude-sonnet-4-6:  1.2k input, 5.3k output, 940.0k cache read, 50.0k cache write ($0.55)

Total cost is an estimate that Claude Code works out on your own computer. The documentation says Claude Code "computes the dollar figure locally from token counts at list price" and adds: "The figure is an estimate, so for authoritative billing see the Usage page in the Claude Console" [1]. The Claude Console is Anthropic's website for accounts that pay per token.

Two groups of people should read this line with care. People on a Pro or Max subscription have usage included in the subscription, and the documentation says the session cost figure "isn't relevant for billing purposes" for them [1]. Organisations with negotiated prices see list price here unless an administrator has set a table of the organisation's prices in a setting named modelPricing. When that table is in effect, the line shows the note at your organization's configured rates. Anthropic's documentation says this setting requires Claude Code v2.1.242 or later [1].

Total duration (API) is the time spent waiting for the model's responses. Total duration (wall) is the time the session has been running [2]. In the sample the first is 6 minutes 20 seconds and the second is 6 hours 33 minutes, so this session was open for hours while the model worked for minutes.

Total code changes counts the lines of code added and removed.

Usage by model is the most useful line. It gives four counts for each model:

• input: new text the model read at full price
• output: text the model wrote, which includes its reasoning
• cache read: text read again from the prompt cache at a lower price
• cache write: text stored in the prompt cache for the first time

The prompt cache is a store, kept by the service that runs the model, of request text it has already processed. Claude Code sends the whole conversation with every request, and the cache lets the service read the unchanged part again at a lower price.

What the sample line shows

It helps to read the sample closely, because it shows where the tokens of that one sample session went. The table below is our own arithmetic. It uses the list prices for Claude Sonnet 4.6 on Anthropic's pricing page, read on 4 October 2026: $3 per million input tokens, $15 per million output tokens, $0.30 per million cache read tokens and $3.75 per million tokens for a five-minute cache write [5].

Count in the sample	Tokens	Share of all tokens	Cost at list price	Share of the cost
Input	1,200	0.1 percent	$0.0036	0.7 percent
Output	5,300	0.5 percent	$0.0795	14.4 percent
Cache read	940,000	94.3 percent	$0.2820	51.0 percent
Cache write	50,000	5.0 percent	$0.1875	33.9 percent
Total	996,500	100 percent	$0.5526	100 percent

The four costs add up to $0.5526, which rounds to the $0.55 in the sample. That match holds when the cache writes are priced at the five-minute rate. The sample does not say which rate applied, so this is our inference.

Three things follow from the table. First, 94.3 percent of the tokens are cache reads, which is the earlier conversation being read again on each request. Second, the text that was new to the model at full price is 1,200 tokens, which is one token in every 830 (996,500 divided by 1,200). Third, the output is half of one percent of the tokens and 14.4 percent of the cost, because output has the highest price.

In a session with these proportions, the way to lower the total is to make the conversation shorter, because new input is 0.1 percent of the tokens. The page on where Claude Code tokens go explains what the conversation contains.

When the totals reset

The documentation says these totals "reset when /clear starts a new session", so the next session's total cost starts at $0. Before Claude Code v2.1.211, the totals kept adding up across /clear for as long as the program stayed open [1]. If you compare two readings, check that no /clear came between them.

The prompt cache line

After the first response, the Session block also shows a line that begins Prompt cache (main). It gives the number of requests, the share of input tokens read from the cache, the number of cache misses, and whether the cache is still usable at this moment. Anthropic's documentation says the line requires Claude Code v2.1.251 or later and covers the main conversation only [1]. The page on how to check your cache hit rate explains each part of that line.

The plan usage breakdown

On a Pro, Max, Team or Enterprise plan, /usage also shows what counts against your plan limits. By Anthropic's documentation, read on 4 October 2026, the breakdown has three parts [1].

Attribution. Recent usage is divided between skills, subagents, plugins and individual MCP servers, each shown as a percentage of the total. A skill is a packaged set of instructions that loads when a task needs it. A subagent is a second copy of Claude that does one task in its own context window. A plugin is an installed package of add-ons. An MCP server is a program that connects Claude Code to an outside tool or data source. The documentation notes that an MCP server's share counts only the requests that used one of its tool results. Before v2.1.222, one call to a server caused every later request to be counted against it, which overstated its share [1].

Behaviour flags. Claude Code names a behaviour, "such as long context or cache misses", when it accounts for 10 percent or more of recent usage, and gives a tip to reduce it [1]. For a long context flag, use the commands on the page about /clear, /compact and /rewind.

Loops. A scheduled task is a prompt that Claude runs again on an interval, for example one started with the /loop command. The breakdown gives a row to each of the scheduled tasks that used the most tokens recently, ordered by total tokens. Each row reports how often the task runs, how many times it ran, its total tokens, its tokens per run, and when it last ran. Claude Code identifies a row by the task's prompt, so a loop that you stop and create again stays in one row. This part requires Claude Code v2.1.242 or later [1].

Press d or w to change between the last 24 hours and the last 7 days. One limit applies to the whole breakdown. The documentation says the figures "are approximate and computed from local session history on this machine", so work done on another device or on claude.ai is left out [1].

If the plan limits fail to load, most often because the usage service is limiting requests, /usage shows the last bars it loaded on this machine within the past 60 minutes, with a note that says how old they are. Press r to try again [1].

/context shows what is in the window now

/usage is a total that grows through the session. /context is a reading of one moment: the text that the next request will contain. Anthropic's command reference, read on 4 October 2026, says it shows current context usage as a coloured grid, with suggestions about tools that use a large share of the context, about memory files that have grown too large, and about a window that is close to full [3].

The breakdown is by category, and Anthropic's documentation says it includes "which CLAUDE.md and auto memory files loaded" [4]. CLAUDE.md is a file of instructions that you write and that Claude reads at the start of every session. Auto memory is a set of notes that Claude writes for itself. Both load before you type, so /context in a new session shows the fixed size that every request will contain.

Two details from the command reference, read on 4 October 2026 [3]:

• When the conversation is larger than the context window, the output includes a warning that says how far over the limit you are and which command frees space.
• In the fullscreen display mode, /context hides the list of single items so that the grid stays visible. Type /context all to show the list.

Use /context at three moments: at the start of a session, to see the fixed size; after a large task, to see what the task added; and before you start a long new task. Anthropic's advice on MCP servers is the same: "Run /context to see what's consuming space." [1] If MCP tools are a large category, read MCP servers or command-line tools.

The status line gives a continuous reading

Both commands need you to stop and ask. The status line shows a reading all the time. It is a bar at the bottom of Claude Code that runs a small script you choose and displays what the script prints. Claude Code can write the script for you. The documentation gives this example [2]:

/statusline show model name and context percentage with a progress bar

Claude Code passes the script a set of named values after each response. By Anthropic's documentation, read on 4 October 2026, four of them matter for token use [2]:

• context_window.used_percentage: the share of the context window in use. It is calculated from input tokens only, which here means new input plus cache writes plus cache reads. Output tokens are left out.
• context_window.context_window_size: the size of the window, which is 200,000 tokens by default and 1,000,000 for models with extended context.
• cost.total_cost_usd: the same estimated session cost that /usage shows, at list price, which "may differ from your actual bill".
• exceeds_200k_tokens: whether the most recent response passed 200,000 tokens in total. This threshold is fixed, whatever the size of the window.

The documentation states that the status line "runs locally and does not consume API tokens" [2]. In plain words, displaying it uses no tokens. Our recommendation is to show the context percentage and the session cost in it at all times.

/insights reports on how you work

/insights answers a different question. The documentation describes it as a report on how you work, where /usage reports how many tokens you used [1]. It reads your recent sessions on this machine and writes a web page that covers what you work on, where requests were misunderstood or code had faults, and suggestions.

Four facts from Anthropic's documentation, read on 4 October 2026 [1]:

• One run analyses up to 200 sessions that it has not seen before.
• The latest report is saved at ~/.claude/usage-data/report.html, and each run also keeps a dated copy.
• The analysis itself uses tokens, and those tokens are counted against your plan or your pay-per-token account.
• Sessions from other devices and from claude.ai are left out.

Reveneau recommends running it once a month. A misunderstood request uses tokens twice: once for the wrong work and once for the correction.

A measuring routine that takes five minutes

Reveneau is an AI software development consultancy in which all code is written by AI, so token use is a running cost of every Reveneau build. Reveneau is independent of Anthropic. This is the routine Reveneau recommends:

1. At the start of a session, run /context and note the fixed size.
2. Keep the context percentage in the status line.
3. After each task, run /usage and read the Usage by model line. If cache reads are large, the conversation is long, and the next unrelated task should start after /clear.
4. Once a week, run /usage, press w, and read the behaviour flags and the attribution shares.

When a number keeps rising and you cannot see why, the page on why usage keeps rising in a long session lists the eight causes that Anthropic documents.