The x20 Receipts
The Follow-Up
Previously: The x20 Is Literal
A few days ago I shared a single day's data: Claude Code API costs versus the Max x20 subscription. One day, roughly $200 in API-equivalent compute, covered by a $200/month subscription. The x20 in the plan name appeared to be literal.
A single day proves nothing. So I ran it for a full week across three Max x20 accounts, working across multiple projects. Same conditions: real work, no stress test, no optimization for or against cost.
The Numbers
Activity
| Metric | Value |
|---|---|
| Sessions | 882 |
| User Turns | 75,817 |
| API Requests | 39,825 |
| Tool Calls | 11,955 |
Tokens
| Metric | Value |
|---|---|
| Input | 2.1M |
| Output | 1.4M |
| Cache Write | 234.9M |
| Cache Read | 3.8B |
| Total | 4.1B |
Cost by Model (API rates)
| Model | Requests | Est. Cost |
|---|---|---|
| Opus 4.6 | 29,734 (74.7%) | $2,573.94 |
| Haiku 4.5 | 8,329 (20.9%) | $127.79 |
| Sonnet 4.6 | 1,360 (3.4%) | $76.72 |
| Sonnet 4.5 | 402 (1.0%) | $48.66 |
| Total | 39,825 | $2,827.12 |
That's one week. Subscription cost: $600/month (3 x $200).
The Math
$2,827 is one week of API-equivalent costs across three accounts. Extrapolated to a month: ~$12,241.
The subscription for those same three accounts: $600/month.
$12,241 / $600 = 20.4.
Per account: ~$942/week, or ~$4,080/month at API rates. The subscription is $200/month.
$4,080 / $200 = 20.4.
The x20 isn't a marketing label. It's what falls out of the arithmetic when you use the tool as a primary coding partner.
Annualized: roughly $49,000 per account at API rates versus $2,400 in subscription fees.
The Cache Story
~5.9 billion tokens per account per month. That sounds dramatic. Look at the composition: the vast majority are cache reads - conversation context that Anthropic's servers recognize from previous turns and skip recomputing, charged at a 90% discount.
Cache reads for Opus 4.6 cost $0.50 per million tokens versus $5 per million for regular input. Without caching, this same week would cost several multiples of $2,827.
The subscription absorbs all of it.
The Model Distribution
74.7% of requests went to Opus 4.6 - the most capable and most expensive model. For complex tasks like architecture decisions, multi-file refactoring, and debugging, Claude Code routes to Opus. For simpler tasks, it falls back to Haiku or Sonnet.
Opus accounts for 91% of the total estimated cost despite handling 75% of requests. For subscription users, this doesn't matter - flat rate. For API users, every Opus call carries premium pricing. That's the gap the subscription fills.
What I Was Actually Doing
Not a benchmark. 882 sessions across three accounts, ten projects. Here's where the tokens went.
The biggest chunk — over a third — went to open-source work. Four OSS projects including puppeteer-capture, lhremote, pcre4j, and one that's still pre-announcement. Feature development, debugging, release prep. The commits are public — go look.
Almost 18% went to Claude improving itself. Custom skills, agents, configuration — Claude Code building its own tooling. The tool that makes the tool better.
Another 18% went to a stealth startup. The rest split between personal brand operations (this very post), and household management — paperwork, recurring tasks, keeping life organized. Nearly 10% of compute went to things no one would call software engineering.
All through Claude Code — if I do something twice, I create a skill/command/agent.
For Individual Developers
| Path | Cost |
|---|---|
| API rates | ~$4,080/month |
| Max x20 subscription | $200/month |
| Break-even | Less than one working day/month |
If you use Claude Code more than a few hours per month at serious intensity, the subscription is the right economic choice. Not close - an order of magnitude.
For CTOs
| Team Size | Weekly API Cost | Monthly Subscription | Annual Savings |
|---|---|---|---|
| 5 engineers | ~$4,710 | $1,000 | ~$233,000 |
| 10 engineers | ~$9,420 | $2,000 | ~$466,000 |
| 20 engineers | ~$18,840 | $4,000 | ~$932,000 |
These assume my usage intensity. Your team will vary. The point isn't precision - it's the order of magnitude.
Caveats
- Sample size of one. All through Claude Code, multiple projects, primarily Opus-tier tasks.
- Pricing will change. AI tooling is in its land-grab phase. These numbers have a shelf life.
- Subscription value depends on usage. Light usage won't hit the 20:1 ratio. This data is for daily, sustained coding.
- Cache rates depend on workflow. Iterative coding with persistent context produces high cache rates. Ad-hoc usage would cost more per token.
The Punchline
The x20 in "Max x20" still isn't a marketing label.
Here are the receipts.