The API-vs-subscription math depends on which cost distribution you're in

You are picking between an API key and a Max seat, and you want a number to decide with. Two numbers are publicly available. They disagree by roughly an order of magnitude, and which one describes you decides the question.

The community claim, sourced from Viberank, is that a working developer's Claude Code usage is worth $1,479/month at API list price — so a $200 seat wins by 5-30x and the only real decision left is which plan's rate limits you can live inside. The official cost guidance says something quite different about the same population.

The two distributions

Read the two middle rows against each other. Anthropic's stated average sits at the price of a Max seat, not a tenth of it. Viberank's median sits six to ten times above it. And Anthropic's 90th percentile — costs.md puts 90% of users below $30 per active day, roughly $600/mo at 20 active days — lands barely above Viberank's 25th, between its p25 and its p50.

Both datasets can be honest at once. They measure different people. Viberank is a public token leaderboard whose participants opted in because they burn tokens; the site gamifies volume. Anthropic's figure comes from enterprise deployments, where a large share of seats belong to people who open Claude Code twice a week. Neither is "the median developer."

What actually follows

The useful move is not adopting either median. It is finding out which distribution you sit in, because your own number is knowable and free.

• If your API-equivalent spend is well under $200/mo, an API key is fine and gives you per-request attribution a seat cannot.
• If it is $400-800/mo, the seat wins on price, but you are close enough to the boundary that the deciding factor is convenience, not arithmetic.
• If it is $1,500/mo or more, price stops being the question. Plan selection becomes a question about limits — the session and weekly windows documented in usage limit errors — and about whether hitting a weekly wall mid-task is acceptable to you.

API billing remains the right answer for cases a seat cannot cover regardless of price: CI, headless automation, per-request cost attribution, and multi-tenant setups where you meter someone else's usage.

Measure your own number, and read it correctly

# Third-party, read-only: sums your local session logs at list price
npx ccusage@latest monthly

Two traps when you compare that number to a seat price:

Second: Claude Code computes every cost figure at list price unless a modelPricing table is in managed settings. If your organization pays contracted rates, the API-equivalent figure overstates what API billing would actually cost you, which biases the comparison toward the seat. When a table is in effect, the Total cost line says at your organization's configured rates. That correction is only available to recent clients: the settings reference says modelPricing "Requires Claude Code v2.1.242 or later", and it is managed-scope only, so below that version every figure you read is list price whatever your contract says.

Also note that /usage and ccusage read local session history on this machine only. Work from another laptop, from claude.ai, or from a cloud session is not in the total, so a single-machine reading is a floor.

What is unverified here

Being direct about the epistemic state, because it is the whole value of this entry:

• The Viberank percentiles are self-reported and self-selected. Nobody has audited them. They are a real measurement of a real population, and that population is unrepresentative by construction.
• Anthropic's $150-250 figure is unsourced in the docs too — no sample size, no date, no methodology. It is an assertion in costs.md, not a published dataset.
• Nothing in the official docs mentions ccusage, endorses it, or publishes a spend distribution you could check either number against.

What would settle it: your own ccusage monthly output, over a normal working month, compared against the price of the seat you are considering. That is a single-sample answer, but it is the only one that describes you rather than a cohort you may not belong to.