Subscription or Pay-Per-Token? I Audited My Own Claude Code Usage to Find Out
Meta description: I ran my real Claude Code usage through live API pricing to see if my $200/month subscription was actually a good deal. The gap was bigger than I expected.
I pay $200 a month for Claude Max 20x, and until recently I had no idea whether that was a great deal or a mediocre one. I'd picked the plan the way most people do: I was already using Claude Code daily, the next tier up looked like "more headroom for more money," and I upgraded without doing the math. That's a strange way to run a decision I make in every other part of the business — I wouldn't greenlight an ad spend or a tool subscription without knowing the return, but I'd been treating my own AI usage as a fixed cost I didn't need to interrogate.
What changed my mind was a simple question I couldn't answer off the top of my head: if I paid for this usage per token instead of a flat monthly fee, would I be paying less, more, or wildly more? I didn't want an estimate. I wanted the real number, built from my own session logs and current API pricing, not a vendor's marketing math or a Reddit thread's back-of-envelope guess.
So I ran the audit. The answer wasn't close, and it taught me more about how to think about subscription-vs-API decisions than I expected.
Why I wanted a real answer, not a guess
The subscription-vs-API question comes up constantly for anyone building with AI tools, and the advice you find online tends to be generic: "subscriptions are better for heavy users, API is better for light or spiky usage." That's directionally true and practically useless, because it doesn't tell you which one you are.
I had a second reason to care. My AI usage has been climbing for months as I've leaned harder into agentic workflows — longer sessions, more parallel subagents, more tool calls per task. A cost that felt reasonable in April could be a completely different number in July. Subscription pricing is flat, so it hides that drift completely. You don't feel the increase until you go looking for it.
The only way to answer this honestly was to pull my actual token consumption and price it at Anthropic's current published rates, then compare that number to what I'm actually paying. Not a forecast. Not an industry average. My own usage, my own bill.
How I ran the actual audit
Claude Code keeps a local record of every session — every request, every token, every model — in JSONL transcript files on disk. I used ccusage, an open-source CLI built specifically to parse those local logs and price them against Anthropic's live per-model rates. Nothing about this pulls from a vendor dashboard or a marketing claim; it reads my own files on my own machine and applies the same published API pricing anyone can look up today.
The output breaks usage down by day and by month, across every model I've touched: Claude Opus 4.8, Claude Sonnet 5, Claude Haiku 4.5, and Claude Fable 5. For each one, it multiplies real input tokens, output tokens, and cache reads/writes by that model's current per-million-token price, and adds it up. That total is the number I'd owe if I were paying per token instead of a flat $200 a month for Max 20x.
I pulled four months of history: April, May, and June in full, plus July through the 20th.
What my real usage showed
April was a false start — I was mostly running other tools that month, with Claude Code as a side project, so it's not a fair baseline. May and June are where things get interesting, because both were full months of real Claude Code usage with no other tool splitting the load.
| Month | API-equivalent cost (real usage) |
|---|---|
| May 2026 | $237.08 |
| June 2026 | $214.91 |
| July 1–20, 2026 | $7,814.13 |
May and June already tell you something: in both months, what I actually consumed would have cost more per-token than the $200 subscription I was paying. Not dramatically more — 7 to 18 percent over — but the subscription was already the cheaper choice before July even started.
Then July happened. Across the first 20 days of the month, my real Claude usage — priced at today's rates — came to $7,814.13, an average of about $391 a day. At that daily pace, a full 31-day month would land around $12,100 in pay-per-token terms. Against my actual $200 bill, that's roughly 60 times more expensive on the API path than on the subscription.
Some individual days in July, taken alone, would have cost more in API terms than my entire month's subscription. One day alone came in at $930 — over four and a half times the monthly Max 20x price, for a single day of work.
Why the gap gets this big
The jump from June's $215 to July's $391-a-day average isn't noise, and it isn't because I suddenly started typing more. It's a function of how the work itself changed shape.
Heavier agentic sessions burn tokens in ways a single chat message never does. Longer-running tasks read and re-read large context windows. Parallel subagents each carry their own token overhead. Premium reasoning models — the ones I reach for when the work actually calls for them — cost five to ten times more per token than the cheaper tiers. None of that shows up as "using AI more" in any way that feels intuitive; it shows up as a much steeper token bill, invisible until you go measure it.
This is the part flat subscriptions genuinely solve well: they decouple your bill from the shape of your workload. I don't have to think about whether a given task is going to spike my usage, because the ceiling is already priced in. That's valuable specifically because the usage pattern is unpredictable and grows with the work, not because the subscription is "cheap" in the abstract.
The decision framework I'd use for any AI subscription
Stripped of the specific numbers, here's the actual decision process, and it applies well beyond Claude Code:
- Pull your real usage, not your impression of it. I would have guessed my July usage was maybe double June's. It was closer to 36 times June's daily rate. Intuition about your own AI consumption is unreliable, especially once agentic workflows are involved.
- Price that real usage at today's published rates, not a number you remember. Model pricing changes, and a stale figure in your head can be off by a wide margin in either direction.
- Compare against what you actually pay, not the plan's list price minus a discount you assume exists. The subscription number is fixed and known; make the API number equally concrete before comparing.
- Re-run this periodically, not once. My own numbers moved from "subscription is 10-ish percent cheaper" in June to "subscription is 60x cheaper" in July. A decision made in June and never revisited would have been directionally right but numerically almost meaningless one month later.
- Watch for the point where your usage pattern, not your monthly total, changes. The trigger for revisiting this wasn't a bigger todo list — it was a shift toward longer, more parallel, more tool-heavy sessions. That's the leading indicator, not the invoice.
For me, the conclusion is unambiguous: Max 20x stays. But the more useful outcome isn't the verdict — it's that I now have a repeatable way to check it again in three months instead of assuming the same answer holds forever. It's the same kind of real-numbers-over-assumptions approach behind the case studies I publish.
If you're trying to make this same call for your own team's AI stack — subscription tiers, API budgets, or a mix across multiple tools — that's exactly the kind of engagement I take on. Contact me and I'll show you the actual receipts, not a guess.
FAQ
How do I check my own Claude Code usage against API pricing?
Claude Code stores per-session token usage locally as JSONL transcripts. A tool like ccusage reads those files directly and prices the tokens at Anthropic's current published API rates, broken down by day, month, and model. That gives you a real "if I paid per token" figure to compare against whatever subscription you're on.
Is a Claude subscription always cheaper than the API for heavy users?
Not automatically, and not by a fixed multiple — it depends entirely on your usage pattern. In my own audit, a subscription that was only modestly cheaper in May and June turned out to be roughly 60 times cheaper in July, once heavier agentic workflows pushed my real token consumption far past where it had been. The comparison has to be run against your actual usage, not assumed from someone else's numbers.
What's the difference between Claude Max 5x and Max 20x?
Both are usage-based tiers above Claude Pro, priced at $100 and $200 a month respectively, with Max 20x providing a substantially larger usage allowance than Max 5x. Which one makes sense depends on your consumption pattern — the same audit approach used here (real usage priced at API rates) is the way to check whether either tier, or pay-per-token API access, fits your actual workload. Anthropic's Help Center has the current plan details.
Why did my AI usage spike so much month over month?
In my case, the spike tracked a shift in how I was working, not just how much: longer agentic sessions, more parallel subagent calls, and heavier use of premium reasoning models, each of which consumes tokens at a different rate than a single short exchange. A monthly total can look stable right up until the underlying workload shape changes, which is why a one-time cost check isn't enough — it needs to be revisited when the way you use the tool changes, not just when the bill does.