How it works
Why a call costs what it costs, and what happens when a seller fails.
Routing
Every model has a list price that heura sets. Each seller offering that model declares a discount below list, from 0% to 80%. For every call the router looks at the healthy sellers for the model and picks the one with the deepest discount.
Billing
The discount reaches you
You are billed list price × (1 − discount) for the seller that served you. A seller at 40% off bills 60% of list. At 80% off it bills 20%.
| Tokens | List price | Seller discount | You pay |
|---|---|---|---|
| 1,000 in + 1,000 out | $2.00 / 1M, both ways | 0% | $0.004000 |
| 1,000 in + 1,000 out | $2.00 / 1M, both ways | 40% | $0.002400 |
| 1,000 in + 1,000 out | $2.00 / 1M, both ways | 80% | $0.000800 |
Input and output tokens can have different prices. The cost of a call is input tokens × input price + output tokens × output price.
Failover
- The router tries the best seller.
- If the call fails, it moves to the next seller. It never retries a seller that already failed for that request.
- A seller that refuses connections is switched off automatically, so later calls do not wait on it.
- If every seller for the model fails, you get an error and you are not charged.
Models without a price
Only models with a list price are served. A call for any other model returns 400 and costs nothing. This stops a seller from listing a model name that nobody has priced.
What a seller never sees or shows
- A seller can serve a model under its own upstream name. Your Usage page always shows the public model name, never the seller's private one.
- Sellers do not receive your wallet address or your balance.