Pricing

Sellers set their own prices. The marketplace routes buyers to the cheapest available seller.

Pricing Modes

Per-Token Pricing

Seller specifies price per 1M tokens for input and output separately (in microdollars).

json
{ "price_input_per_1m": 12.00, "price_output_per_1m": 48.00 }

Cost Multiplier

Seller specifies a fraction of the model's reference price. Useful for "I'll undercut the market by X%."

json
{ "pricing_mode": "cost_multiplier", "cost_multiplier": 0.50 }

0.50 = buyer pays 50% of the reference price. Reference prices come from PROVIDER_PRICES in the codebase.

Cost Calculation

For each request:

base_cost = (input_tokens / 1M × price_input_per_1M)
          + (output_tokens / 1M × price_output_per_1M)
buyer_cost = base_cost × fee_multiplier / 10000 + flat_fee
seller_earns = base_cost

Fee Multiplier

Stored on-chain in the Settlement contract in basis points. Transparent and auditable.

MultiplierMeaning
10000 (default)1.0x — buyer pays exactly the seller's price, no fee
103001.03x — 3% fee, surplus goes to feeRecipient

Minimum value: 10000 (1.0x). Sub-1x subsidy mode is not supported in SettlementV2.

Price Threshold

Buyers can require a minimum discount below the go-direct reference price:

bash
-H "X-Min-Discount: 30"

The same floor can be set on the API key itself (min_discount_pct in the key's scope) or expressed in the URL (/min30/v1/...). When more than one is present the strictest applies — a per-request value can tighten the key's floor but never loosen it. A request that no offer satisfies fails with minimum_discount_not_met rather than routing to a worse price.

Price ceilings are retired. max_price_per_1m and the X-Max-Price-Per-1M header capped a per-1M input rate, which ignored output and cache rates and told you nothing about whether the price was actually good. Discount is measured against the reference price you would pay going direct, which is the number that matters.

Requests skip sellers whose estimated cost exceeds this threshold.

Provider Pricing Discovery

When sellers connect their API key, the marketplace auto-detects pricing from the provider's /models endpoint. Four parsing layers handle all known provider formats:

  1. Venice format — model_spec.pricing with nested {usd, diem}
  2. OpenRouter format — pricing.prompt / pricing.completion as per-token strings
  3. Generic format — pricing.input / pricing.output as numbers ($/1M tokens)
  4. Reference fallback — static reference prices for known models when provider returns none

Next

Settlement