What does Claude Code normally cost per month?

Light users often stay within Pro/Max; heavy Agent users commonly see $120–$250/month. Anthropic docs cite roughly $150–$250 for active enterprise developers.

How do I check Claude Code usage?

Use /usage in-session; /usage-credits for monthly caps on Pro/Max; export 30-day CSV from Anthropic Console Usage.

What should I cut first when the bill spikes?

Cut wasted turns first: Opus on small tasks, never /clear, full reruns after failure, missing .claudeignore. Downgrade models second.

I audited 30 days of Claude Code billing—
and got a shock

June 18, 2026 · Claude Code / AI cost · ~6 min read

Developer terminal and Claude Code billing dashboard: 30-day usage audit — I thought Claude Max at $100/month was the ceiling—until I stacked Console exports, session logs, and credit-card line items.

May 19–Jun 18, Console CSV + credit-card reconciliation: $247.18 total. I thought Claude Max capped at $100; actual overage was $147.18; 62% of the overage came from 5 long sessions (40+ tool-call rounds). Everything below is tables—numbers you can reproduce.

1. 30-day total

Item	Amount	Notes
Claude Max subscription	$100.00	Fixed monthly
API overage	$147.18	Long sessions + Opus + large input
Total	$247.18	$8.24/day avg.
Cursor overage same period (reference)	$87.40	Dual-tool stack, not in table above

After this table: I stared at the “API overage $147.18” row for a few seconds—I’d already bought the top tier, so why nearly double the subscription on top? What stung more: Cursor had another $87 not counted here. Two cards, two “very capable interns” I’m apparently raising at once.

2. Weekly curve

Week	Dates	Daily avg.	Main driver	Week total
Week 1	5/19 – 5/25	$4.2	Short Q&A, learning the tool	$29.4
Week 2	5/26 – 6/1	$6.8	Agent CI fixes, first 38-round session	$47.6
Week 3	6/2 – 6/8	$9.1	Two repos in parallel, no /clear	$63.7
Week 4	6/9 – 6/18	$11.8	Xcode 26 beta + full-repo reads	$106.5

The curve gave me chills: workload only grew maybe 30%, but daily spend climbed from $4.2 to $11.8. In week 4 I kept telling myself I “wasn’t doing much”—Console doesn’t care. One session with 1.8M input tokens is literally a dinner out.

3. Task types (breaking down the $147.18 overage)

Task type	Sessions	Share	Typical rounds	Main issue
CI / test green	7	28%	25–55	Full rerun after failure
Cross-file refactor	5	22%	30–48	Default Opus, no file list first
Explore unfamiliar repo	9	19%	15–35	No .claudeignore, scanned Pods/logs
Blog / docs writing	6	16%	20–40	Many polish passes
Parallel subagent	2	12%	40+ ×2	Same repo read twice
Other	—	3%	<10	—

Bittersweet: the easy “small fixes” were only 3%; the CI green work I dread ate 28%. What scared me wasn’t Claude Code being expensive—it was me dumping the dirty work on it with Opus as default.

4. Top 5 most expensive sessions

#	Date	Cost	Rounds	Cause
1	6/12	$41.2	52	Fix main CI + read full 12MB beta log
2	6/3	$33.7	47	18-file API change, Opus, no /clear
3	5/29	$28.4	41	Two subagents in parallel
4	6/7	$22.1	36	No ignore file, scanned vendor
5	6/15	$18.9	34	Blog multi-pass polish

I read these five rows over and over: five sessions ate 98% of the overage; #1 at $41.2 was my own prompt—“help me get main CI green.” No session boundaries, too lazy to /clear, everything on Opus—I built this bill myself.

5. Hidden multipliers (not listed separately in CSV)

Multiplier	Typical outcome	What I hit
Failed rerun	Lid closed → task interrupted → read disk from scratch	+40%–70% tokens; two runs ~$19
Context snowball	No /clear—every round bills full history	Single session 8k → 940k input
Dual-tool stack	Claude Code + Cursor each burn L1	Cursor extra $87/mo

These rows aren’t in the CSV, but the card remembers: two lid-close reruns were pure habit and cost ~$19 for nothing. Context snowballed from 8k to 940k and I didn’t notice—until reconciliation showed how expensive “try again” really is.

6. Cost rules (target: $120/month)

Rule	Practice	Effect (measured)
Model routing	Sonnet daily; Opus only for architecture / big refactors; then `/clear`	Opus usage ↓ ~60%
Session boundaries	One subtask per session; ban “while we’re at it…”	Fewer long sessions
.claudeignore	Exclude node_modules, Pods, DerivedData, *.log	Same prompt: 1.2M → 90k input
Quota monitoring	`/usage`, `/usage-credits`; Console alerts at 80%/100%	No end-of-month shock

Two days in, half the anxiety gone: since 6/16 I’ve followed this table—daily avg. dropped from $11.8 to $6.3. The win that felt best was .claudeignore—one file edit, same prompt, input down an order of magnitude. Better than swapping models.

Common commands

/model sonnet    # default
/model opus      # complex tasks only
/clear           # required after each subtask
/usage           # current session estimate
/usage-credits   # set monthly cap

7. FAQ

Question	Answer
Why over $140 on top of Max $100?	Opus long sessions, parallel Agents, huge input burn included credits fast; overage bills at API rates.
What’s normal per month?	Light: within subscription; heavy Agent: often $120–$250 (official reference).
Cut models or seats first?	Cut wasted turns first (long sessions, no ignore, reruns), then consider model downgrade.
Paying twice with Cursor?	Overlap exists; keep one primary Agent shell—see IDE layering article.

If you just got scared by your bill: don’t rush to drop seats or quit—ask: Opus on small stuff? One session for everything? Agent reading all of node_modules? You’ll probably save the first chunk there.

8. Verdict

After 30 days of reconciliation, my biggest takeaway isn’t “learn to save money”—it’s accepting an Agent isn’t a $20 Copilot. It’s a partner billed by the step. Bill = steps × context × unit price, plus rerun multipliers; $247 shocked me because it shattered my “$100 cap” fantasy. Export CSV on a weekend, mark your Top 5, write three rules of your own—worth doing once. ROI still positive for me, but only if you read the bill before you talk about addiction.