What Claude Code Costs for a Team of 20

Problem: Every published break-even table answers the pricing question for one person. You are buying for twenty. You multiply, get a number, take it to the CFO, and it is wrong in a way that will not show up until month four.

Quick Win: The arithmetic is easy. At Anthropic's published Team prices, twenty Premium seats (a seat is one paid license for one person) cost $2,000 a month, $24,000 a year on annual billing. Twenty Standard seats cost $400 a month, $4,800 a year. But two costs are missing from that line, and above roughly fifteen people they get bigger than the license: the seats nobody uses, and the engineer hours spent keeping the shared setup working. Budget for those two before you argue about which plan.

The Per-Person Answer, And Why It Breaks At Twenty

The standard analysis, including our own break-even math on plans versus pay-as-you-go, asks one question: does this person use enough to justify a flat monthly fee instead of paying per use? Right question for one person. It compares a fixed cost to a variable one and picks the smaller.

At twenty people, two things become true that were not true at one.

First, you are no longer buying for a known user. You are buying for a spread. Some people run the tool all day. Some open it twice a month. You pay the same fixed price for both.

Second, you now have a shared setup: shared rules, shared permissions, shared connections to your other systems. At one person, the owner is the user and the cost is invisible because it is the same hour. At twenty, the owner is a specific engineer and the cost is real and unbudgeted.

So the honest twenty-person question is not "which plan." It is: which recurring work is worth running unattended at all, and who owns the thing that runs it. Unattended means the AI runs it on a schedule with nobody watching.

The Multiplication Table: What Twenty Seats Actually Costs

Start with what is published, so nobody has to trust an estimate. These are Anthropic's own listed prices as of this writing (claude.com/pricing).

Two things to read from this before anything else.

Paying monthly instead of annually costs $6,000 a year on twenty Premium seats. A 25% premium for the privilege of not committing. It is the cheapest saving available and the one most often missed, because whoever clicks "subscribe" is rarely whoever owns the budget.

The gap between the cheap seat and the expensive seat is 5x. $24,000 a year versus $4,800. That $19,200 gap is the entire decision, and it should be made per person, not for the whole team at once. Anthropic's pricing page currently lists Claude Code across all its paid plans, and the more expensive seat is the one sold on higher usage limits. What each tier includes has changed more than once in the last year, so confirm it on the day you buy rather than trusting any table, including this one.

On top of the seat, Team and Enterprise plans can enable extra usage billed at standard pay-per-use rates, meaning Anthropic's published price list rather than a flat fee. Spend caps are set by the account owner at both the organization and individual level. Hit the cap and extra usage pauses until the next billing period (Anthropic support). Leave those caps unset and you have a variable line item in a budget you presented as fixed.

What A $100 Seat Has To Earn

A token is the unit AI billing is measured in, roughly three quarters of a word. Published pay-per-use rates for Claude Opus 5 are $5 per million input tokens and $25 per million output tokens. Content the tool has already read once can be reused at 10% of the input rate (Anthropic pricing docs).

Take a mid-size working session as 100,000 tokens in and 20,000 tokens out.

So a $100 seat pays for itself somewhere between 100 and 156 mid-size sessions a month, which across roughly 21 working days is about 5 to 7 sessions a day. That is a genuinely heavy user, not someone who opens it when stuck.

Multiply by twenty and the team number is 2,000 to 3,100 sessions a month to justify $24,000 a year, roughly 95 to 150 sessions a day across twenty people.

Write that down. It is the only honest test of whether the seats are earning, and notice what it is measured in: sessions, not seats. Nobody buys sessions, so nobody tracks them, which is exactly how the next cost hides.

Omitted Cost One: Seats Bought And Not Used

Software license waste is well measured and it is not small. Zylo's 2026 SaaS Management Index, built on more than 40 million licenses and $75 billion in spend under management, found that organizations leave an average of 36% of their software licenses unused (Zylo). The same report puts median software spend at $9,455 per employee per year and notes AI application spending up 108% year over year.

We will not invent a usage rate for your team. Here is the full range instead, so you can find your own row.

At the industry average, the $100 seat is really a $156 seat. At 50% usage it is a $200 seat, double the price you thought you agreed to.

This is not a licensing problem. It is a measurement problem. The invoice tells you how many seats you bought. It does not tell you how many produced work. Two numbers, and only the wrong one is on the bill.

Omitted Cost Two: The Person Who Maintains The Shared Setup

Here is the line no per-seat comparison contains, because it never arrives as an invoice. Somebody writes the shared rules file. Somebody decides which tools the AI may touch and which files it may not. Somebody fixes it when a connection to another system changes. Somebody answers "why did it do that" for nineteen other people. Somebody notices when output quality drifts.

That work is constant, and the evidence for why is unambiguous. In Stack Overflow's 2025 developer survey, the single biggest frustration, cited by 66% of developers, was AI output that is "almost right, but not quite," and 45% said debugging AI-generated code takes more time than expected. More developers now actively distrust the accuracy of AI output (46%) than trust it (33%) (Stack Overflow). Google's 2025 DORA report, an annual survey of software teams based on nearly 5,000 technology professionals, found 90% using AI at work and 30% reporting little or no trust in the code it generates (Google).

Almost-right output does not fail loudly. It gets reviewed, corrected, and quietly absorbed into somebody's week. Multiply by twenty and it stops being a rounding error. We wrote up what that upkeep contains in the reliability iceberg: the build takes a weekend, keeping it running takes months.

The Break Point: Three Hours A Week

Do not guess at this. Calculate it. Two inputs, both of which you already know.

Monthly license cost  =  seats x seat price
Monthly upkeep cost   =  hours of upkeep x fully loaded hourly cost

Fully loaded means salary plus taxes, benefits and overhead, not the number on the offer letter. At twenty Premium seats on annual billing the license line is $2,000 a month. Here is the crossover at three common hourly costs.

At $150 an hour, three hours a week of unowned upkeep costs the same as the entire license bill for twenty people. Three hours a week is not a project. It is one person answering questions and patching a settings file on Friday afternoon. It never gets logged, it never gets approved, and it is half your true spend.

Note what does not appear in the crossover: headcount does not drive upkeep. The number of distinct setups does. Ten people sharing one well-maintained configuration generate less upkeep than five people with five configurations each. That is why the decision changes shape somewhere around fifteen to twenty people. Not because the per-use price goes up, but because that is where "one person's setup" quietly becomes "everyone's setup with no owner."

If the license bill and the maintenance bill are the same size and only one of them has a name on it, you do not have a pricing question. You have an ownership question.

What To Standardize Before You Buy Twenty Of Anything

The most useful finding in the 2025 DORA report is not about tools at all. It is that AI acts as an amplifier: "AI doesn't fix a team; it amplifies what's already there. Strong teams use AI to become even better and more efficient. Struggling teams will find that AI only highlights and intensifies their existing problems" (Google Cloud).

The report is blunt about where the return comes from: "not from the AI tools themselves, but from a strategic focus on the quality of internal platforms, the clarity of workflows, and the alignment of teams." It reports a direct correlation between the quality of a company's shared internal setup and its ability to get value from AI.

Read that as a purchasing rule. Buying twenty seats into an unstandardized environment buys twenty slightly different versions of the same mistakes, faster.

Three things are worth fixing first, and all three are cheap:

1. One shared rulebook. One file the AI reads every session that encodes how your company actually works. We wrote the pattern up in the company rulebook.
2. A named owner. One person accountable for the shared setup, with the hours in their plan rather than stolen from it. The crossover table above is the argument for funding that role.
3. A short list of recurring work worth running unattended. Proposals, quotes, recurring reports, weekly briefs. That list is the actual return on the license spend, and it is what we build as department automation.

Where This Goes Wrong

You buy one tier for everyone. The 5x price gap between Standard and Premium exists because usage varies at least that much across a team. One tier for twenty people means you either underserve the heavy users or overpay for the light ones. Usually both.

You measure seats and call it adoption. Seats are what you bought. Sessions are what you got. If your only report is the invoice, you find the gap at renewal, which is the most expensive moment to find it.

You leave the extra-usage caps unset. Extra usage bills at the published pay-per-use rates. Without deliberate caps at the organization and individual level, a fixed budget becomes a variable one and you learn about it in arrears.

You treat a plan as a reliability guarantee. It is not. It is access. Whether output is trustworthy and safe to act on lives in the shared setup, the unbudgeted line. Two thirds of developers already name "almost right" output as their top frustration. A more expensive seat does not move that number.

You buy the seats before the standards. The common one, because seats are procurable in an afternoon and standards are not. It is also the one DORA's data argues against most directly.

When The Honest Answer Is Fewer Seats

Sometimes the right recommendation for a twenty-person team is eight expensive seats, twelve cheap ones, and one funded owner.

That configuration costs $9,600 a year in Premium seats plus $2,880 in Standard seats, so $12,480 against $24,000 for twenty Premium. The $11,520 difference funds roughly 77 to 115 hours of a named owner's time at $150 to $100 an hour, which is enough to build the shared rulebook and keep it working for a year.

Then you expand on evidence. Anthropic's admin controls let you manage seat allocation, set spending limits at the organization and individual user level, and view Claude Code usage analytics (Anthropic), so after one quarter you know who crossed the 5-to-7-sessions-a-day line. Upgrade those people. Leave the rest.

Narrow, then evidence, then expand. It costs less and tells you more than buying twenty seats and finding out in month eleven that Zylo's 36% applied to you too.

Related Reading

• I Can Build It. Should I?, the true-cost calculator for the build side of the same decision
• The Reliability Iceberg, what the maintenance line actually contains
• The Company Rulebook, the shared file that makes twenty seats behave like one standard
• What a 1,000-Agent Workflow Costs, the same accounting applied to unattended work instead of seats

Frequently Asked Questions

Is the expensive seat worth it for a whole engineering team?

Only for the people who clear the usage bar. At published pay-per-use rates, a $100 seat needs roughly 5 to 7 mid-size sessions a working day to beat paying per use. Some of your team clears that easily. Most teams have a tail that does not. Buy per person, not per team.

Does the annual discount matter at twenty people?

More than most expect. Twenty Premium seats cost $24,000 a year on annual billing and $30,000 billed monthly (claude.com/pricing). That is $6,000 for flexibility you probably will not use, on a tool your team is going to keep.

What single number should we track after buying?

Sessions per active person per month, checked against the 100-to-156 break-even. It is the only figure that tells you whether a seat is producing anything. The invoice cannot, and neither can a headcount.

If your twenty-seat budget is really a two-line budget with only one line written down, the fix is not a different plan. It is deciding which recurring work is worth running unattended, standardizing it once, and giving it an owner so the maintenance stops being invisible. That is the part we install. See what we build for companies, or write to hugues@topr.io.