The setup
The skill is weckr-margin-audit, one of the four free Weckr Claude Skills. It encodes the method from the unlimited plan heavy tail math: cost a typical and a heavy user per plan at current prices, margin against plan price, flag the tail. To test it we opened a clean Claude Code session, from an empty directory so nothing could leak in, and asked:
My SaaS has one Pro plan at 49 dollars a month, unlimited AI usage
on claude-sonnet-4-6. A typical user makes 1200 calls a month, about
1500 input and 500 output tokens each. A heavy user makes 6000 calls.
Do I lose money on anyone?What Claude answered
Condensed to the numbers, which are verbatim:
Cost per call: 1500/1e6 x $3 + 500/1e6 x $15 = $0.012
typical 1,200 calls $14.40/mo margin +$34.60 (71%)
heavy 6,000 calls $72.00/mo margin -$23.00 (underwater)
break even ~4,083 calls/month
"Yes, you lose money on heavy users. Any user past ~4,080 calls
a month costs more in AI than they pay. One 6,000 call user wipes
out the AI margin of ~1.5 typical users. With unlimited flat
pricing, your risk is entirely in the tail."It closed with the fix ladder, output caps, caching, cheaper model routing, metered overage, per user caps, and, unprompted, the caveat that this is AI cost margin only, before infrastructure, support, or fees. Every number checks out by hand against the current Sonnet 4.6 prices of $3 and $15 per million tokens.
Why this needed a skill at all
Two things fail without one, and they are exactly the two that matter:
- Prices. A base model quotes model prices from training data, and providers move faster than training cutoffs. OpenAI cut a model 80 percent in July, covered in the GPT-5.6 price drop. The skill pins math to the live LLM pricing feed with a dated fallback table, which is why the $3 and $15 in the transcript are right.
- Method. Asked casually, assistants audit the average user, and the average user is precisely where unprofitable plans hide. The skill hard requires costing a heavy user, because the tail decides flat pricing, not the mean.
Encoding know-how so an assistant applies it reliably is the whole idea of skills; the wider comparison with live data connections is in MCP server vs Claude Skill.
Where the audit honestly stops
The transcript costs the users we described. Your actual distribution contains users nobody described: the integration that retries in a loop, the team member with a script, the customer whose workflow is 40 times median. An estimated audit detects that the plan design is exposed to the tail. Naming which accounts are past the line this month takes production data, per user, against plan price, which is the measuring problem described in tracking AI costs per user and the reason Weckr exists.
Run it on your plans
/plugin marketplace add Ghiles3232/weckr-sdks
/plugin install weckr@weckr
"Which of my plans lose money on AI cost? Free at $0,
Starter at $19, Pro at $49 unlimited, all on gpt-5.4-mini..."Free, no account, and the skill is markdown you can read before trusting. If you prefer sliders to prompts, the break even calculator covers the single plan case.
FAQ
Can Claude audit whether my AI pricing plans are profitable?
Yes, if you give it the method and current prices, which is what the free weckr-margin-audit skill does. You describe your plans and rough usage, and Claude costs a typical and a heavy user per plan at current model prices, computes margins, finds the break even call count, and flags plans where the heavy tail goes underwater. It shows every step so you can challenge the assumptions.
What did the audit find in your test?
We described a $49 unlimited plan on Claude Sonnet 4.6 with a typical user at 1,200 calls a month and a heavy one at 6,000. Claude computed $0.012 per call, a 71 percent margin on the typical user, a $23 monthly loss on the heavy user, and a break even at roughly 4,080 calls, then noted the result covered model spend only, before infrastructure or support costs.
How is this different from asking plain Claude the same question?
Two failure modes go away. Plain Claude prices models from training data, which is reliably stale, and it tends to audit the average user, which is exactly the user that hides the problem. The skill pins prices to a live feed and forces the heavy tail into the analysis, because flat pricing is broken by the tail, not the mean.
How do I run this audit myself?
Install the free plugin inside Claude Code with /plugin marketplace add Ghiles3232/weckr-sdks then /plugin install weckr@weckr, and ask which of my plans lose money on AI cost, with your plan prices and a rough sense of usage. The skill does the rest and shows its work. No account, no signup.
What are the limits of an estimated audit?
It costs the users you describe, not the users you have. Real usage is a distribution with surprises in it, and the genuinely unprofitable accounts are usually ones nobody guessed. Treat the audit as the risk detector for the plan design, then measure real per user cost in production to name the actual accounts.
Keep reading
The audit finds the risk. Production names the users.
When the estimated audit says your tail is exposed, the next question is who is in it right now. Weckr answers that continuously: real cost per user per feature against what each user pays, with the underwater accounts flagged as they happen. See exactly that screen on the live demo, or start from the AI cost and margin guide.