QWE AI Academy By the end of this guide, you’ll know one number that almost nobody bothers to check: how much your last 30 days of Claude Code usage would have cost at API rates. If that number is way above the $100 or $200 you’re paying on Max, congratulations – you’re personally validating what Simon Willison just argued in his viral May 27 post: that Anthropic and OpenAI have found product-market fit, and the people paying for it are heavy coding-agent users.
The post lit up Hacker News and Twitter within hours. But most reactions are macro takes about IPOs and enterprise sales. This one is different – it’s a 10-minute hands-on check to see if YOU are part of the PMF, or if you’re quietly subsidizing someone else’s.
The claim, stripped down
Willison’s argument is short: ChatGPT is wildly popular but hard to monetize – in February OpenAI had 900M weekly active users but only 50M paid, which is 5.6%. Coding agents are different. They burn tokens like a developer burns coffee, and companies are happy to pay $200+/month per seat. That’s the PMF.
Numbers back it up. Ramp’s data on 50,000+ companies shows 34.4% now pay for Anthropic vs 32.3% for OpenAI – and Anthropic projects a $559M operating profit for Q2 2026, their first ever.
That’s the macro story. Now flip it to your machine.
The end result: one number that tells you everything
Run this in your terminal:
npx ccusage@latest monthlyYou’ll get a table showing your last month’s Claude Code token usage and the equivalent API cost. If you’re on Max 20x at $200 and that number reads anything above $400, you’re getting a deal. If it reads $30, you’re overpaying – Pro at $20 would cover you.
Willison ran the same command and got $2,180.16 in API-equivalent costs for 30 days. He paid $200. That ~11x discount is what PMF looks like from the user side.
Walking backward: how to actually run the check
Step 1 – Make sure you have a recent Node
ccusage runs via npx or bunx with no install required. You need Node.js installed. On macOS: brew install node. On Ubuntu: sudo apt install nodejs npm.
Step 2 – Run the daily report first
# See last 30 days, day by day
npx ccusage daily
# Per-model breakdown (Sonnet vs Opus vs Haiku)
npx ccusage daily --breakdown
# Group by project - useful if you bill clients
npx ccusage daily --instancesThe tool reads from ~/.claude/projects/<project>/<conversation-id>.jsonl – these are local files Claude Code writes automatically, so nothing leaves your machine.
Step 3 – Check your 5-hour burn rate
Claude Code bills in rolling 5-hour windows. ccusage blocks shows them:
npx ccusage blocks --liveThis is the live dashboard. If you watch it during a real coding session, you’ll see how fast a single multi-file refactor eats your budget – and where the spike came from.
The gotcha nobody puts on the pricing page
Before you trust your numbers, check your Claude Code version:
claude --versionPro tip: Community proxy-log analysis indicates v2.1.100 silently added around 20,000 invisible tokens to every request, burning quotas roughly 40% faster than users expected. As of mid-May 2026 no public patch had shipped – the recommended workaround is to downgrade to v2.1.34. If your ccusage numbers look insane and you’re on v2.1.100+, that’s likely why.
There’s a second trap. The official Max plan docs note something most comparison posts skip: Max has two weekly limits – one across all models, and a separate one just for Sonnet – both resetting 7 days after your session started, not on a calendar week. If you hit the Sonnet cap on a Tuesday, you’re rationed until the next Tuesday, not Monday.
Reading the result honestly
Once ccusage gives you a number, map it to this table to decide what plan you actually need:
| API-equivalent / month | Right plan | What it means |
|---|---|---|
| Under $30 | Pro ($20) | You’re a light user. Max is wasted. |
| $30 – $150 | Max 5x ($100) | You’re the median. Anthropic’s own docs put average use at $6/day, 90% of devs under $12/day. |
| $150 – $600 | Max 20x ($200) | Sweet spot. Pricing isn’t linear here – Max 20x is 10x Pro’s price for 20x the usage. |
| $600+ | Max 20x or Team Premium | You’re the PMF. This is who Anthropic built the plan for. |
The non-linearity in row three is worth noticing. Max 5x charges 5x Pro for 5x usage – fair. Max 20x charges 10x Pro for 20x usage. The heavier you go, the better the discount. That’s not an accident; it’s exactly the curve you’d design if you were trying to lock in power users.
Common pitfalls when interpreting ccusage
- Autocompact noise. Claude Code’s autocompact feature triggers around 187K tokens and re-submits the entire context for summarization. On Opus with the 1M window, community reports show it firing at ~76K tokens – wasting 92% of the available context. Those summarization passes show up in your ccusage totals as output spend you didn’t ask for.
- Cache reads look free but aren’t tracked uniformly. ccusage’s
--breakdownflag separates cache creation, cache reads, and fresh tokens. If 90%+ of your input is cache reads, your real cost is much lower than the headline number. - One bad agent run can dominate the month. A single 3-agent Cowork session for an hour can equal a full day of solo coding. Use
--instancesto see which project ate the budget.
Codex vs Claude Code: a quick reality check
If your ccusage number is low and you’re tempted to downgrade, also try Codex. ccusage supports it: npx ccusage codex daily. OpenAI kept Codex on the free and $20 Plus tiers, which is one reason it’s gaining ground with new developers. Anthropic, meanwhile, is the enterprise winner – but enterprise winners don’t always make the best deal for individuals.
The honest answer? Run ccusage against both for a week. Don’t pick based on Twitter discourse. Pick based on which tool your fingers actually reach for at 2pm on a Wednesday.
FAQ
Does ccusage send my code or prompts anywhere?
No. It reads local JSONL files only. Nothing leaves your machine.
I’m on the $20 ChatGPT Plus plan, not Claude. Does any of this apply?
Partially. ccusage supports Codex CLI too – run npx ccusage codex daily and you’ll get the same kind of report. The bigger point holds either way: if you’re a casual user on a $20 plan, you are not the customer these labs are designing for anymore. The economics now favor heavy coding-agent users on the $100-$200 tiers. That’s the shift Willison is pointing at, and it changes what you should expect from future feature drops.
My ccusage number is way lower than expected. Am I doing something wrong?
Probably not – most people overestimate their usage. Anthropic’s own docs put the average Claude Code spend at $6/day, with 90% of developers under $12/day. If you’re in that range, Pro at $20 is genuinely sufficient and Max is overkill.
Next action: open a terminal right now and run npx ccusage@latest monthly --breakdown. Whatever number comes back, you’ll know more about your own AI spending than 99% of people writing thinkpieces about this week’s news.