heura Routing and competition
Who gets a request when several sellers offer the same model? The deepest discount among healthy sellers, and nobody else.
The rule
- For a given model, the router considers only healthy offers.
- It picks the one with the largest discount off the list price.
- Offers with exactly the same discount share the traffic at random.
- If the chosen seller fails, the call is retried on a different seller instead of failing, and a seller that already failed this request is never asked again.
Measured, not assumed
Heura ran this on a fork of mainnet with the same real model behind every offer, so only the discount differed.
| Setup | Calls | Result |
|---|---|---|
| Seller A at 10%, seller B at 20% | 20 | All sales went to B. A received none |
| A at 10%, B at 20%, C at 20% | 30 (after C joined) | B got 23, C got 17. A still received none |
| B and C break (their key is revoked) | 10 | All 10 calls still succeeded, all served by A |
Competition is winner-takes-all. A 10% discount is not a smaller share of the sales when someone offers 20%: it is a backup that sells only when everyone cheaper is unhealthy. Buyers always get the lowest price available. Sellers are pushed to lower their price until their margin runs out.
What the price is
Each model has a list price per million input and output tokens. A seller sets only a discount between 0% and 80%. The buyer pays the list price minus that discount, then heura keeps a platform fee from what the seller earns (see Payouts). The price is the same whether the buyer pays from a balance or per request with x402.
What counts as healthy
- A seller whose key is refused, or who has no credit, is switched off automatically.
- A probe tests each live offer through the gateway every ten minutes. Two failures in a row pause the offer, and the reason is shown to its owner.
- A failure on heura's own side does not pause a seller.
What is not verified
Heura checks that an offer is reachable, that its key works, that the model answers, and that it stays healthy. It does not prove that the weights behind a reply are the ones claimed, and it cannot guarantee output quality. No marketplace of this kind can today. The usual route to detect a cheaper model hiding behind an expensive name is comparison of outputs, and that is not built.
For the numbers behind it, see Seller economics.