Skill cost is not a uniform thing. A pricing analysis on three numbers costs cents. A market research pass across 12 competitors costs dollars. Pretending the costs are the same is how AI subscriptions go broke.
Skill OS classifies every skill into one of four cost classes. The classification is visible. The per-tier daily caps are visible. Here is how the system works.
Four classes
- Light. Median cost per fire: $0.01 to $0.05. Most catalog skills. Includes the entire advisor stack (Pricing Surgeon, Growth Machine, Coach, Stoic).
- Medium. Median: $0.10 to $0.30. Skills that read multiple inputs, run multiple LLM passes, or compose multiple outputs. Editorial Calendar, MCP Wrapper Builder, Code Reviewer Pattern.
- Heavy. Median: $0.50 to $1.50. Skills that crawl the web, generate images, render documents, or call expensive third-party APIs. PDF ebook designer (with cover art), Market Researcher, Slide Deck Builder with image gen.
- Premium. Median: $2.00 to $10.00. Skills that produce large deliverables (full 60-page proposals, multi-hour video transcription + analysis, batch image gen). Rare in the catalog.
Each skill declares its cost class in the manifest. The router reads it. The user sees it on the catalog page.
Per-tier daily caps
Each tier has a current daily spend cap, in addition to message-count caps. These figures are provisional while the cost dashboard accumulates production data; the published pricing page is the source of truth.
- Free: roughly $5/day. Light skills only. No medium, no heavy, no premium.
- Pro ($99/month): roughly $50/day. Light unlimited. Medium up to 40/day. Heavy up to 5/day. No premium.
- Max ($199/month): roughly $150/day. Light unlimited. Medium unlimited. Heavy up to 20/day. Premium up to 3/day.
- Ultra ($299/month): roughly $500/day. Everything unlimited within the cap.
The cap is the cap. When you hit it, skill firing pauses for the day. Light replies still work (they cost cents). Medium and above gate out. The user sees the explanation on the next attempt.
Credit packs
The cap is a default, not a hard ceiling. If you hit it, you can buy a credit pack:
- $10 for 100 messages
- $40 for 500 messages
- $75 for 1,000 messages
Credits never expire. They auto-tap when your monthly cap is exhausted. Stripe payment links. One-tap purchase. No subscription change required.
We did not invent this. Most modern AI products that take cost seriously have a similar surface. We adopted it because the alternative (no cap at all) is the surface that ends in surprise four-digit invoices.
Why this matters more than it looks
Cost-class-aware routing is the thing that makes a 200+ skill catalog economically possible. Without it, every skill fires at the same priority, and a single user can run a $200 skill 50 times in a session and bankrupt the model layer.
We learned this the hard way during early load testing. A user who got into a feedback loop with the Market Researcher skill ran it 23 times in 8 minutes. The cost was $73. The user did not realize until they checked the dashboard. We added the daily cap and the per-skill warning the next week. The same loop today is gated at fire 6 with a confirmation card.
Where cost matters less than it looks
For 95 percent of users, 95 percent of fires are light. The advisor stack is light. Most content drafting is light. Most code review is light. The cost system is invisible to most users most of the time.
The cost system becomes visible when:
- You use heavy skills regularly (market research weekly, deck building monthly).
- You hit a cost-class transition (running 6 heavy fires in a day on Pro tier).
- You build a workflow that chains skills (one user message fires 4 skills in sequence; the total can cross a threshold).
In those cases, the system tells you. It does not silently ship a bigger bill.
How we classify in practice
Three signals decide cost class:
- LLM cost. Token volume + model tier. A skill that uses Sonnet for one short pass is light. A skill that uses Opus across four long passes is medium or heavy.
- Third-party API cost. A skill that calls a web crawler with a per-page fee is medium minimum. A skill that calls an image-gen API is heavy.
- Output volume. A skill that produces a one-paragraph reply is light. A skill that produces a 60-page deliverable is premium.
We re-audit classification quarterly. When a skill's median cost drifts (because the underlying model got cheaper, or because we tightened the prompt), we re-class it. The catalog page reflects the current class.
What we do not do
We do not gate access to skill descriptions or skill documentation by cost class. The full catalog is visible to every user, free or paid. You can see what each skill does even if you cannot run it at your tier.
We also do not hide the cost. Every reply that fires a non-light skill includes a short cost note ("This used a Heavy-class skill. Today's heavy-budget remaining: 4 fires.") The transparency is the contract.
What you can do
If you want to be cost-aware: nothing. The system protects you by default.
If you want to use the catalog aggressively: upgrade to Max or Ultra. The daily caps are designed to let you actually use the catalog without thinking about per-skill cost.
If you hit a cap mid-month and need to keep going: buy a credit pack. No commitment, no upgrade dance, no waiting.
The pricing page has the full breakdown. The principle is one line: cost is real, cost is visible, and the cap is the protection.