How to · Margin

Is Your Unlimited AI Plan Profitable? Do the Heavy Tail Math

By Ghiles Asmani, founder of Weckr · Published August 5, 2026

Short version: an unlimited AI plan is judged by its heaviest user, not its average one. Compute your cost per call, find the break even call count for the plan price, then cost a heavy user honestly. In the example below, a $49 unlimited plan holds 71 percent margin on the typical user and loses $23 a month on the heavy one, and one heavy user cancels the margin of about one and a half typical users.

Flat pricing feels safe because the revenue is predictable. The cost is not, and unlimited plans attract exactly the users who make it least predictable. This is the ten minute audit that tells you whether yours is fine, fragile, or already underwater.

The worked example

One plan: Pro, $49 a month, unlimited AI, on claude-sonnet-4-6 ($3 input, $15 output per million tokens). Calls average 1,500 tokens in and 500 out:

cost_per_call = 1500/1e6 * $3 + 500/1e6 * $15
              = $0.0045 + $0.0075 = $0.012

Now cost three real shapes of user:

typical   1,200 calls * $0.012 = $14.40  ->  margin $34.60  (71%)
heavy     6,000 calls * $0.012 = $72.00  ->  margin -$23.00 (underwater)
break even: $49 / $0.012 = ~4,083 calls

Every call past 4,083 in a month is a call you pay for. The heavy user is not exotic: 6,000 calls is 200 a day, which one integrated workflow or one enthusiastic team member produces easily. And this is AI cost margin only, before infrastructure, support, or payment fees, so the real line is lower than 4,083.

Why your average says everything is fine

AI usage is skewed. If nine users make 1,200 calls and one makes 6,000, average usage is 1,680 calls, average cost $20.16, average margin a reassuring 59 percent. Meanwhile the tenth user loses you $23 every month, invisible inside that average. The mean is the wrong lens; the distribution is the right one. This failure mode, one account quietly underwater while the dashboard smiles, is the subject of when one user costs more than they pay.

Worse, unlimited plans select for the tail. The users who most value unlimited are the ones who intend to use the most, so the plan design recruits its own worst case. The pricing model tradeoffs are covered in flat subscriptions for AI SaaS.

The fix ladder, cheapest friction first

  • Cap output tokens. Output is 5x input on Sonnet; a tighter max_tokens trims every call invisibly.
  • Cache repeated context. A large system prompt at the cached rate is a tenth of the price on Anthropic.
  • Route heavy traffic to a cheaper model. claude-haiku-4-5 at $1 and $5 cuts the example cost per call to $0.004, moving break even past 12,000 calls. See model downgrade on budget.
  • Include a quota, meter above it. Keeps the headline simple and puts a price on the tail instead of a subsidy.
  • Per user spending caps as backstop. Block or auto downgrade past a monthly limit, so no single account can run away. Mechanics in per user spending caps for LLM apps.
  • Raise the price. Last resort, but $49 covering a $72 user is not a pricing strategy, it is a donation.

Audit yours in ten minutes

Swap in your own numbers: plan price, model prices from the live LLM pricing feed, your token shape, and honest call counts for a typical and heavy user. Or delegate it: the margin audit skill in the free Weckr Claude Skills runs this exact method when you ask Claude Code which of your plans lose money, and the break even calculator does the single plan version in your browser.

FAQ

Is an unlimited AI plan ever profitable?

Usually yes for the typical user and no for the heavy tail, and the tail decides. In our worked example a $49 unlimited plan on Claude Sonnet 4.6 keeps a 71 percent AI margin on the typical user but loses $23 a month on a heavy user. Whether the plan is profitable overall depends entirely on how many heavy users you attract, which unlimited pricing actively selects for.

How do I find the break even usage on my plan?

Divide the plan price by your cost per call. At $0.012 per call on a $49 plan, break even is about 4,080 calls a month. Every call past that is paid for by you, not the customer. Knowing this one number changes the conversation from is unlimited safe to how many users are past the line.

Why do averages hide unprofitable users?

Because AI usage is heavily skewed. Most users are light, a few are extreme, and the average lands in a comfortable looking middle. A plan can show healthy average margin while its top five users each lose you money every month. The distribution matters, not the mean, and one heavy user can cancel the margin of several typical ones.

What are the fixes if my unlimited plan is underwater?

Roughly in order of customer friction: cap output tokens, cache repeated context, route heavy users to a cheaper model, add an included quota with metered overage above it, set per user spending caps that block or downgrade past a limit, or raise the price. Most products need only the first three plus a cap as a backstop.

How do I audit my plans without building anything?

Two free options. The margin audit skill in the Weckr Claude Skills lets you ask Claude Code which of my plans lose money on AI cost, and it costs a typical and heavy user per plan and flags the underwater ones. Or use the break even calculator at useweckr.com/calculator with your own plan price and model. Both are estimates; production data settles it.

Keep reading

Why Flat Pricing Breaks for AI

Why flat plans break, and the three fixes that hold.

One User Costing More Than They Pay

How to spot a customer who costs more than their plan.

We Let Claude Audit Our Plan Margins

The transcript: break even found, heavy tail flagged.

Estimates find the risk. Production finds the users.

This audit tells you whether the plan design is exposed. It cannot name the specific accounts past the line this month, because that takes per user cost data from production. Weckr tracks real cost against plan price for every user and flags the underwater ones as they happen, not at the invoice. See the flagged users on the live demo, or start from the AI cost and margin guide.

See the dashboard with real data, no signup needed.

Try the demo →