Why we do not mark up inference
Most API resellers charge 20–40% on top of the official model price and call it a platform fee. We take the opposite approach: zero markup on tokens, a small fee on top-up.
When you buy credits from an API reseller, you are usually paying two things at once: the model vendor's price, and a hidden margin. The margin is invisible because it is baked into the per-token rate, which makes cost forecasting nearly impossible as models change.
The problem with token markup
Markup on tokens punishes exactly the behaviour you want to encourage. Teams that use cheaper models to save money still pay a percentage premium on every call. Teams that optimise prompts still pay the same percentage. The margin scales with your success, not with the service you receive.
It also makes comparison shopping dishonest. A gateway advertising "GPT-5 at $8 per million input tokens" is not comparable to one advertising the official $1.25 — unless you already know the list price, which most buyers do not.
What we do instead
Inference is billed at upstream list price, to the micro-dollar. Our revenue is a platform fee charged once, when you top up: 4% below $100, tapering to 2.5% above $2,000. After the top-up, a token costs exactly what the vendor charges.
| Top-up | Platform fee | Effective markup on usage |
|---|---|---|
| $20 | 4% | 0% |
| $100 | 3.5% | 0% |
| $500 | 3% | 0% |
| $2,000 | 2.5% | 0% |
Why this is better for both sides
You get predictable unit economics: your cost per request is the vendor price plus nothing. We get a business that does not depend on you overpaying for tokens, and a pricing page that stays honest when a vendor cuts prices at 2am.
When an upstream provider lowers prices, our users see the reduction on their next request without a migration, a plan change or a support ticket.