What It Costs to Run an AI App on SkillSafe

The short answer: a run costs the model’s list price for the tokens it actually used, plus two equal cuts on top — 10% to the platform and, by default, 10% to the app’s creator — charged in credits at $1 = 10,000 credits, rounded up to whole credits with a 1-credit minimum. A chat turn on the default Workers AI model costs 16 credits ($0.0016); a hefty Claude Sonnet 5 run costs a few hundred. This post walks through every number, because opaque AI pricing is a choice and we made the other one.

Credits: $1 = 10,000

Credits are not a token or a points scheme — they’re dollars at a fixed denomination. One credit is $0.0001, chosen so that genuinely cheap runs can be billed exactly instead of rounding up to a whole cent. Your balance is shown in dollars, activity rows show exact costs, and cash amounts convert at the same rate in both directions.

What a Run Costs: One Base, Two Equal Cuts

Every model an app can run has a published rate on the model catalog — dollars per million input and output tokens. The arithmetic on top of it is deliberately boring:

base      = provider list cost for the tokens used + $0.001 per-job overhead
platform  = 10% of base
creator   = the app's markup (0–100%, default 10%) of the same base
you pay   = base + platform + creator, rounded up to whole credits

Both cuts are percentages of the same base. That matters for two reasons. It means the platform’s 10% is its entire revenue on a run — there are no seat fees, listing fees, or a second cut taken out of the creator’s side. And it means an app advertising “+10%” is charging 10% of compute cost, not 10% of some already-marked-up figure.

At the default markup, the creator and the platform earn the same amount from a run. That symmetry is the point, and it’s a change from how this used to work: until July 2026 the platform billed a 1.2× multiplier on compute and then took a 10% fee out of the creator’s markup, which added up to roughly twice what the creator made. Both of those are gone.

The mechanics protect you on both sides:

• Before a run, the platform estimates the worst case and places a hold — you see the estimate before clicking run.
• The run settles at actual usage reported by the provider; the unused hold is refunded. Failed runs refund fully.
• Each run rounds up to a whole credit, minimum 1.
• If your balance can’t cover the worst case, the run scales its output cap down to your budget instead of failing.

Worked Examples

Real arithmetic at current catalog rates, for a run of typical chat size, on an app at the default 10% creator markup:

Scenario	Model	Tokens in / out	Compute	You pay	In credits	Creator earns
Chat turn, default model	Gemma 4 26B (Workers AI)	1,500 / 400	$0.00027	~$0.0016	16	$0.000127
Mid-size run, budget model	GPT-5 mini	3,000 / 1,000	$0.00275	~$0.0045	45	$0.000375
Premium run	Claude Sonnet 5	3,000 / 1,000	$0.02400	~$0.0300	300	$0.002500

All rows include the per-job overhead and whole-credit rounding, at the default 10% creator markup. Two things that table shows plainly. Model choice changes what a run costs by ~19× at identical token counts, which is why apps declare their model on their detail page. And on the cheapest models the $0.001 overhead is most of the price — that Gemma turn is $0.00027 of inference inside a $0.00127 base. The overhead pays for the work every run does regardless of model (queueing, the hold-and-settle ledger entries, storing the result), and it is what lets both 10% cuts be worth a whole credit instead of rounding away. If you are optimising for cost per run, a cheap model stops helping below about a cent.

The Creator Margin

App creators may set a markup of 0% to 100% of a run’s base cost. It’s shown on each app’s detail page (“Usage-based + N% creator margin”, with the effective per-token rate), and it’s how a skill maker earns from their work. Creators keep 100% of it — nothing is deducted. On the Sonnet 5 row above, a creator at the 100% ceiling would price that run at 525 credits ($0.0525) and keep $0.025 of it.

Creators can move that number on a slider in their app dashboard and watch what each setting pays them per run against their app’s own average run cost. The creator-side view of this math is in How to Monetize a Claude Code Skill.

Running for Free: Signup Credit, Daily Grants, Free Apps

• $0.30 signup credit for every new verified account.
• $0.20 daily grant on each day you actively use the platform.
• Free grants expire after 7 days; purchased credits and earnings never expire.
• Guests can run guest-enabled apps without an account at all.
• Apps labeled “Completely free” in the directory are verified from source to never charge usage credits, and creators can sponsor usage from their own wallet with daily caps.

At default-model prices, the signup grant alone is roughly 185 chat turns — the free tier is real.

BYOK: Flat 1-Credit Runs on the Publisher’s Own Key

Apps whose publisher attached their own Anthropic or OpenAI API key bill differently: the publisher’s key pays for inference, markup is forced to zero, and you’re charged a flat 1 credit per run as pure platform overhead. It’s the cheapest possible lane and shows up like any other run in your activity.

Holds, Refunds, and Budget-Scaled Runs

Because settlement is capped at the pre-run hold, an app can never charge more than the estimate you saw. Since July 2026, hold/refund pairs display as a single activity row with the exact settled cost, and under-funded runs execute budget-scaled rather than throwing a 402 — both in the changelog.

Frequently Asked Questions

Why is the billed rate above the provider’s list price?

The 10% platform margin absorbs payment-processing fees, failed-run refunds, and hosting, so there are no seat fees, listing fees, or surprise charges anywhere else. The creator’s markup is a separate, always-displayed cut of the same base.

How do I see what a specific app will charge me?

The app’s detail page shows its pricing lane (free / usage-based / margin %) and its model with the effective per-token rate. Before any run, the app shows the hold estimate.

What happens if a run fails?

The hold is refunded in full. Settlement only ever charges provider-reported usage on completed runs.

Can I buy credits without a subscription?

Yes — one-off top-ups from $5 to $1,000 on your billing page, always in whole cents. There is no platform seat fee for running apps; plans and team pricing are on the pricing page.

Conclusion

The whole cost model fits in one sentence: tokens used at list price, plus 10% to the platform and the creator’s markup on the same base, in credits at $1 = 10,000, rounded up, min 1, capped at the hold you saw. Every input to that sentence is public — the model catalog has the rates, app detail pages have the margins, and your activity ledger has the receipts. If you’d rather be on the earning side of it, start with turning your skill into an app.