The Anthropic bill arrives 3 weeks after you incur it. By then, the debugging session that burned $40 in a single afternoon is a distant memory — and the team member who ran it has moved on to a different project.
We learned this the hard way. A developer on our team racked up $180 in Claude Code API costs over two days. Nobody noticed until the monthly invoice landed. The session logs showed 2.1M output tokens — roughly 14x what we expected. Turned out a subagent was stuck in a retry loop, regenerating the same file diff 40 times.
You can't optimize what you can't see. Here are four ways to track Claude Code token usage before the bill surprises you.
1. The Built-in: /usage
Claude Code ships with a /usage slash command. Run it inside any session and you get a token breakdown for that session:
Fast. No setup. But it only covers the current session — yesterday's sessions are gone, and there's no team-wide aggregation. For subscription (Pro/Max) users, the dollar figure doesn't appear because usage is baked into the plan.
For API users, /usage is your baseline sanity check. Run it before closing a session. If the cost looks wrong, dig into the token breakdown — a low cache hit rate or unusually high output tokens usually points to a problem.
2. The Source of Truth: Anthropic Console
The Usage page in the Anthropic Console is the billing source of truth. It shows daily token consumption broken down by model, with costs calculated at official rates.
The catch: data lags by 24-48 hours, and the granularity stops at the organization level. You can see that 2M Sonnet tokens were consumed on Tuesday, but not which developer or which project burned them.
For teams with shared API keys — which is most teams — the Console tells you that costs spiked, not why. You need session-level data for that.
3. The Community Tool: ccusage
ccusage is a local CLI that reads Claude Code's session log files and produces detailed breakdowns:
Output looks like this:
| Date | Model | Input Tokens | Cache Read | Cache Write | Output Tokens | Cost |
|---|---|---|---|---|---|---|
| 2026-07-01 | sonnet-4-6 | 1.2M | 8.4M | 0.3M | 0.4M | $12.18 |
| 2026-07-02 | sonnet-4-6 | 0.8M | 6.1M | 0.1M | 0.3M | $8.42 |
| 2026-07-03 | opus-4-8 | 0.4M | 2.1M | 0.2M | 0.1M | $6.80 |
What makes ccusage better than /usage: it separates cache reads from cache writes from uncached input. That distinction matters because cached reads cost $0.30/M on Sonnet while uncached input costs $3/M — a 10x difference. If your cache read ratio drops, your bill spikes, and the aggregate numbers from the Console won't tell you why.
ccusage also exports JSON, which makes it scriptable for dashboards and alerts.
4. The Team Solution: OpenTelemetry
For teams running multiple Claude Code instances, OpenTelemetry export is the way to get per-developer attribution. Claude Code natively exports metrics (token counts, costs, tool activity) and events via the standard OTel protocol.
Set the environment variables:
Point it at any OTel-compatible backend — Grafana, Datadog, Honeycomb — and you get per-session, per-developer, per-model dashboards. This is what we run internally. The OTEL_LOG_TOOL_DETAILS flag is worth enabling: it logs which tools are consuming the most tokens, which is how we caught our agent's tool output waste problem.
What to Actually Look For
Once you have tracking in place, three metrics matter more than the rest:
Cache hit rate. Below 60% on Sonnet means you're paying 10x more for input than necessary. Check our caching guide for fixes.
Output token ratio. If output exceeds 15-20% of total tokens, a subagent is probably stuck in a retry loop or generating excessive diffs. That's the pattern that cost us $180.
Session cost variance. Track the 90th percentile session cost, not just the average. A few runaway sessions will dominate your bill, and the average hides them.
We built Synrouter partly because we got tired of discovering costs weeks after the fact. The gateway logs every request with full token attribution — cached vs uncached, per session, per developer — so you see costs as they happen, not when the invoice arrives.
If your Claude Code bill has been creeping up and you can't pinpoint why, sign up to get started — per-session cost visibility is included from day one.
Read next: Claude Code API Pricing: Real Costs & How to Cut Your Bill 85%