Skip to content
Agentic Control Plane

Claude Code Cost Tracking: Token Counters Show What You Spent, Not Where It Went

David Crowe David Crowe · · Updated · 6 min read
claude-code cost optimization agents
Share X HN LinkedIn

These figures are from our June 2026 snapshot (210,840 calls). A refreshed July corpus (285,814 calls) shows the same pattern more sharply, with 97% of spend in the orchestration loop. See the updated teardown.

One full working day of Claude Code: 276 model calls and 1,697 tool calls, priced at $148.16 at API rates. On a Max subscription your invoice says nothing like that, which is why “what does Claude Code actually cost?” has produced a small crop of trackers. Most of them are token counters. A counter tells you what you spent. What changes behavior is where it went. For that session, the answer was that all of it was loop cost. Not a dollar went to the model producing answers. All of it was the loop re-reading 72 million tokens of accumulated context, turn after turn. The last 28 turns of the day bought 10% of the output.

The ACP Cost page for one week: $1,353 spent with the change against the week before, sessions, per-session average and cache hit, and spend by day stacked by model (Fable, Sonnet, Opus) with the prior week as a dashed line

Update, September 2026: the Cost page above is what the console opens on now. It shows the week against the week before, spend by day by model, and the levers to pull. The session X-ray below is still where you break down a single run.

Session cost X-ray of a real working day: $472.03 at API rates, 1,052 loop turns, 100% loop tax at a 100% cache hit rate, the cumulative bill curve, and 3,744 free tool calls beneath it, with 4,790 allowed, 6 denied, 115 PII findings
curl -sf https://agenticcontrolplane.com/install.sh | bash
claude   # model turns estimated from the transcript at API rates; session X-ray in the console

With plain claude, the ACP hook reads the session transcript and sends each turn’s token usage to the gateway. The gateway prices it at list rates and labels it an API-rate estimate, not a charge. For metered cost, launch with claude-acp instead. It routes model calls through the ACP proxy, and on exit it prints the session total with a link to the breakdown:

ACP · session: 276 model calls · 1,697 tool calls · $148.16 @ API rates
→ https://cloud.agenticcontrolplane.com/sessions/…

You can’t see any of that in a total. This post maps the Claude Code cost-tracking options, what each can and can’t tell you, and what per-turn attribution adds.

What token counters do well

There are three kinds, and the first two are useful.

Built-in /cost. Claude Code shows the current session’s token usage on demand. No setup, but it covers one session and is gone when the session ends.

Local counters like ccusage. These read Claude Code’s local session logs and roll them up: daily totals, per-model splits, how much of your subscription you’re using. If your question is “am I getting my money’s worth from Max?”, a log-reading counter answers it in one command. No proxy, no account, no config. Use one.

LLM proxies. Route Claude Code’s API traffic through a proxy (the LiteLLM-style setup) and you get metered totals per key or per user. It’s the team-level version of the same answer.

All three add up tokens and multiply by price. Their unit is the session or the day, and that’s where they stop.

What a total can’t tell you

The $148 session looks the same, as a total, as a session where the money bought 276 useful answers. A counter can’t tell them apart. The expensive part of an agent is the structure, not any one call. The loop re-reads the whole growing transcript on every turn, so the same tokens get billed again and again. Across 210,840 metered tool calls on our own workspaces in June, that loop was 89% of total spend, and one frontier model on the loop drove 80% of the bill from 7.6% of the calls. We call it the loop tax. A sum hides it.

A token counter can’t answer these:

  • Loop or leaf? Was the spend the model deciding what to do next and re-reading context, or doing the sub-task that produced output?
  • Which turn? Where did cost speed up, and did the last N turns buy anything? (Ours bought 10%.)
  • Is caching working? The $148 day ran at a 100% cache hit rate, which is the only reason it wasn’t roughly 10 times more. That’s a per-call property, and a total hides both the win and the miss.
  • Which action? 1,697 tool calls happened. Which of them set off the expensive re-reads?

None of this is exotic. They’re the questions you’d ask about any bill: not just how much, but for what.

Per-turn attribution

The alternative is to tag every model call with what it was: which session, which turn, which model, loop or leaf, cache hit or miss, and what that one call cost in dollars. That works even on a subscription, where the price you actually pay per call is $0. With claude-acp, Anthropic bills you exactly as before, because your own login passes through. The pricing is a view, not a charge.

Then the session ends with a line instead of a mystery:

ACP · session: 276 model calls · 1,697 tool calls · $148.16 @ API rates
→ https://cloud.agenticcontrolplane.com/sessions/…

The X-ray behind it shows the cost curve turn by turn. That’s where the levers show up, because the levers that move agent cost are about structure. Put a cheaper model on the loop (in our June data the frontier model on the loop cost about 114 times per call what the cheap model did). Check that the cache is actually hitting. Stop the runaway session at turn 30 instead of turn 300. A daily total can’t tell you any of that.

For a subscription user, the same data answers a different question. $148.16 at API rates against a flat monthly fee settles the seats-versus-API decision with your own numbers instead of a forum thread.

What to use

For the flag-by-flag version of what Claude Code shows natively and the config that meters every call, see the Claude Code cost tracking reference.

  • To know what you spent, use /cost or ccusage. They’re free, local and good.
  • To know what your team spent, use a metering proxy with per-key totals.
  • To know where it went and what to change (loop versus leaf, per turn, priced even on subscription), use ACP’s cost X-ray. It comes with the same install that checks your tool calls:
curl -sf https://agenticcontrolplane.com/install.sh | bash

Plain claude keeps its policy-checked hooks and gets estimated cost from the transcript. claude-acp is the same session with model calls routed through the proxy, so cost is metered and the proxy can also redact tool results. If your question is which turn, which model and which loop drove the bill, that’s what AI agent cost monitoring is for.


Cost figures are from our own sessions and workspaces (June and July 2026 snapshots). They are our own use, not customer data.

Running Claude Code? One command adds a hook. Each tool call is recorded, the session's model calls are priced, and a force-push, destructive SQL or curl | sh waits for your OK in the session.

macOS · Linux · WSL
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
Windows · PowerShell
irm https://agenticcontrolplane.com/install.ps1 | iex

What the installer writes  ·  Claude Code guide  ·  free up to 5 agents

Share X HN LinkedIn
Get the next data drop
What agents actually cost, new tool-surface captures, and the occasional incident post-mortem — sent when we publish something worth your inbox, not on a schedule. Unsubscribe anytime.
Share: Twitter LinkedIn
Related posts

← back to blog