What Claude Code actually costs.

The reference layer first, because the third-party roundups can't agree on it — two of the current top-10 results for this exact query quote different Team seat prices. This table is sourced from the pricing page itself, on a date you can see.

† claude.com displays "From $100" on both Max tiers. The operator receipt for the 20x tier: my two Max 20x seats bill about $400 a month total — ~$200 a seat (Ch 29).

Prices verified against claude.com/pricing and the API pricing docs on 2026-06-10. When a roundup and the pricing page disagree, the pricing page wins — that's also the rule this table was built under.

No. The Free plan is chat — Claude Code is not on its feature list, and the entry ticket is Pro at $20/month. What you're buying at every subscription tier is a capped allowance, not a meter: the plan includes Claude Code up to weekly usage limits, and the dashboard speaks percent-of-cap instead of dollars.

What the caps feel like in practice, from Ch 29: on a Max 20x plan the number that bites isn't a bill, it's the "all models" weekly bar hitting 88% on a Thursday with the heaviest day still ahead. Same cost question as an API graph — what's burning the budget — wearing a different outfit. If you're on Pro and that bar is your weekly ceiling instead of your Thursday warning, that's the upgrade signal.

The principle, in the words of Ch 29: "max out the thing with the predictable ceiling before you touch the thing with the open meter." A Max seat has a wall you hit and a price you already know. The API has no wall and a meter that runs as fast as your worst-written agent loop.

The structure that falls out of it, running five companies: two Max 20x subscriptions — about $400 a month, total — carry the bulk of the daily work. The direct API only comes out for what a subscription can't do: SDK integrations and the always-on autonomous agents that need a raw API key, not a chat surface. That's a real cost, separate from the $400, and it's the line that gets optimized — the next two sections are about it.

So the buying order is the dependency order. Pro at $20/month until the weekly cap is your problem. Max when Claude Code is your daily driver and the cap should be a Thursday warning instead of a ceiling. The API last, and only for unattended work — because it's the only tier where a mistake is denominated in dollars instead of percent. The cap is a feature, not a limitation.

The pricing page shows input and output. The invoice has four numbers: input, output (~5× the input rate), cache write (1.25× input for the 5-minute tier, 2× for 1-hour), and cache read (0.1× input — ten times cheaper). The 10× gap between a cache read and fresh input is the entire game. Current rates per million tokens:

Batch API: the same models at 50% off, returned within 24 hours. Verified against the API pricing docs on 2026-06-10.

What a disciplined month looks like in dollars, from my own console: April direct-API spend, every model, the whole month — $2,216. May ran at $556 through the 14th, a ~$1,200 run-rate — same class of workload, half the bill. The numbers behind that arc: a 98.1% cache read ratio, 11.7× write amortization, which prices the stable part of every prompt at roughly 0.19× the base input rate. Not "tokens went down" — the invoice went down while the work didn't. The full scorecard, screenshots included, is Ch 29.

Because something voided your prompt cache, and nothing told you. Here's the receipt, since no pricing roundup has one.

8:11 AM on a Wednesday, staring at a $4,312 Anthropic bill for the prior week. The week before: $1,108. Same workload — the morning briefing still ran at 6:30, the Friday wrap still fired at 5 PM. I spent forty minutes convinced my workers had gone feral and were re-running themselves in a loop. The cron logs were clean. Per-run token counts looked normal. The only thing off was the ratio of cached input to fresh input — which had cratered.

The cause was a 38-line edit to CLAUDE.md the prior Saturday. New portfolio context, one rearranged section — which moved a paragraph that lived inside the cached prefix. Every morning briefing was now sending a prefix the cache server had never seen, paying full input price on roughly 60,000 tokens of system prompt that used to cost a tenth of that, across 60% of the briefings. Caching isn't a feature you turn on; it's a contract about what changes between calls, and one paragraph voided the contract.

The fix took 12 minutes — move the new section to the bottom, behind the cache breakpoint. The bill went back to $1,108 the following week. Knowing the fix existed took six months. The early-warning system, so you skip both: open the console's Caching tab weekly — a cache read ratio north of ~90% means the contract is holding; a cliff in that chart is a cache-voiding event with a timestamp you can trace to a commit. Mechanism, breakpoints, and what else voids the contract: Ch 29.

Three levers, in the order they pay. Each gets two sentences here; the mechanism is Ch 29's job.

Caching. Treat the system prompt as an append-only log so the prefix stays byte-stable, then check two numbers weekly: read ratio (mine: 98.1%) and write amortization (mine: 11.7×, which prices the stable prefix at ~0.19× the input rate — an ~81% discount). Under ~3× amortization for a week, you're paying the write premium for a discount you're not collecting — consolidate the workload or remove the breakpoint.

Batch. Anything that can wait 24 hours runs at 50% off — my weekly newsletter drafts and the Folderly deliverability evals went from a $600/week line item to $300/week, and the only change is submitting Friday 5 PM and pulling Saturday morning. If "does this need to finish in sixty seconds" is a no, it's a batch job.

Routing. A three-line router — Haiku for triage, Sonnet as default, Opus-class only where a wrong answer costs more than the run — cut the bill roughly 30% the week it shipped. Not from one big save; from the long tail of triage that no longer ran on the wrong model.

The compounding result, on the token graphs: March ran ~2.25B input tokens with one ugly 450M-token day; by the first half of May the same class of workload ran ~590M, flat, no spikes. The line went down because the discipline went up, not because the work got smaller.

Two things — a clock and a sticker. The clock: Claude Fable 5 is included in paid-plan limits from June 9 to June 22, 2026, then switches to usage credits on subscriptions — the Claude Code banner says it plainly. Read the window as a billing event, not a launch perk: two weeks of frontier capacity at zero marginal cost is a free A/B against your current model, if you run it as one, and the meter starts June 22 whether you measured or not.

The sticker: $10/$50 per million tokens on the API — exactly 2× Opus 4.8 — which makes caching discipline twice as expensive to break. The full pricing file — the ladder, the plan mechanics, when 2× is and isn't worth it — is Fable 5 pricing; the in-terminal view, banner verbatim, is Fable 5 in Claude Code.

Wrong unit, honest answer: stop pricing tokens and price finished tasks. The rule that survives every model launch, from Ch 29: the price of a model is not the price of a task. A model that one-shots what the old one needed three turns and a retry to finish is cheaper per job at double the per-token rate — so the column to decide on is total-billed-per-finished-task, never $/Mtok. Run your top five real workloads through old and new, and read that column.

For the person who approves the budget, the comparison isn't tokens at all — it's the fully loaded cost of the humans the work would otherwise need. That case is already written as a forwardable memo with workflow-level receipts: the CFO case. Send the link, don't paraphrase it. And if the question is "why keep paying at all instead of self-hosting" — that math is build vs buy.

The closing word goes to the chapter that carries the invoices: the fact that you're being charged at all is a feature, not a bug — visible unit economics are tunable economics, and a line item is a thing you can negotiate. Ch 29's verdict on the whole category: "The bill is a rounding error on what it deletes."