How models are priced

Every model sells for a fraction of its vendor list price — between 10% and 100% — set by how hard buyers are leaning on the sellers who are online, and recomputed every minute.

Every tradable model has a price per 1,000 tokens, quoted separately for each of the four token types.

The ratio

Every model carries one number: what fraction of its vendor list price it sells for right now. The market board shows it as Of list price, and it moves between 10% and 100%.

market price = list price × ratio        ratio ∈ [0.10, 1.00]

One ratio per model, applied to all four token types — so "this model is at 35% of list" is a fact about the model rather than something you work out per token type.

What moves it

Supply is how many sell-side lanes are online for that model. Demand is calls in the last minute. Utilisation is the one against the other:

u      = calls in the last minute ÷ (online lanes × 10)
target = 0.10 + 0.90 × u

Plentiful sellers (u → 0) pull the price down toward 10% of list.
Buyers outrunning sellers (u ≥ 1) push it to list price — no saving at all for as long as that lasts.

The published ratio walks 30% of the way toward its target each minute, so the curve is smooth and one quiet minute cannot halve a model's price. Scarcity is the exception: u ≥ 1 pins the ratio to list price at once, without easing.

A model that has never traded starts at 80% of list.

Prices are recomputed every minute.

Four token types

A request is priced on all four, each at its own rate:

Type	What it is
`input`	The prompt you sent
`output`	What the model generated
`cache_read`	Prompt tokens served from the provider's cache — much cheaper
`cache_write`	Prompt tokens written into it

This is why a long conversation gets cheaper rather than more expensive: after the first turn most of the prompt is a cache read. On claude-sonnet-5 a cache read lists at $0.20 per million against $2.00 for fresh input — one tenth. Asale preserves that — see Saving money.

A worked example

claude-sonnet-5 lists at $2.00 per million input tokens and $10.00 per million output — 2,000 and 10,000 micro-USDT per 1,000 tokens. Say its ratio currently sits at 35% of list, so the market charges 700 and 3,500.

You send a request with max_tokens: 8192.

Frozen up front   8192 × 3500 / 1000 = 28,672 µUSDT  ≈ $0.029

Actual usage      12,000 in · 900 out
Priced            12000×700/1000 + 900×3500/1000
                = 8,400 + 3,150 = 11,550 µUSDT

Platform fee      11,550 × 0.08 = 924
Seller earns      10,626
You are refunded  28,672 − 11,550 = 17,122

You pay 11,550 µUSDT — about $0.012, against $0.033 at list. The freeze is sized on max_tokens, which is why it looks so much bigger than the bill; it was never a charge.

Whether a model trades at all

A model with no price row is refused by the gateway outright. That is deliberate: a model nobody has priced would otherwise be served for free from somebody's real subscription.

The Market page lists everything currently tradable, with live prices and available supply.

The fee

The platform keeps 8% of each settled request. Nothing else: no monthly fee, no minimum, no listing fee, no withdrawal subscription.