Basics · 2026-11-12 · 321AI engineering notes · get a key
Quick answer: Multi-supplier bidding is the mechanism behind low prices on an aggregation platform: every model is quoted by several independent suppliers at once, their prices and quality metrics are published side by side, and routing automatically prefers the cheapest healthy lane. Competition does the price discovery — you just read the board.
Popular models are served by more than one operator. Each operator runs its own delivery lane — its own capacity, its own upstream contracts, its own cost structure — and quotes what that lane costs to run. The marketplace aggregates every quote for every model into one board: the model catalog shows, per model, the market floor price, how many suppliers quote it, the average stability and P50 latency. On a busy model that board holds thousands of live quotes across the whole market.
Depth varies by model, and that variation is information. A model with dozens of quoting lanes is a commodity — price competition is fierce, floors sit low, and failover is deep. A model with two or three lanes is thinner: still functional, but the success-rate column matters more and the floor moves further when one lane has a bad day. Reading depth before volume is the habit that separates a marketplace user from someone who just copied a model name.
Each quote carries more than a price. Reading all four fields together is what turns a price table into a decision:
| Field | What it tells you | How to act on it |
|---|---|---|
| Market floor price | The cheapest healthy quote right now | Budget planning and per-task cost math |
| Supplier count | How many lanes quote the model | More lanes = deeper automatic failover |
| Success rate | Recent share of calls that completed | Prefer high lanes for user-facing traffic |
| P50 latency | Typical response time | Interactive apps care; batch jobs less |
A lane cannot hold a premium for long while a cheaper healthy lane is visible on the same board. The routing engine scores lanes in milliseconds — context length, task profile, live health, price — and picks the cheapest fast healthy one. Conditional routes re-evaluate within 5 minutes as the market moves, so a lane that raises prices or degrades quietly loses traffic without anyone filing a ticket. The floor you see on the pricing page is the visible result of that pressure.
Floors are not constant. They drift with lane health (a quarantined lane stops quoting), with supplier pricing changes, and with demand-side pressure at peak hours. What stays constant is the mechanism: the board is public, re-selection happens automatically within 5 minutes, and the floor you see is always the cheapest currently-healthy quote — not an average, not a promise, not a locked rate. For budgeting, use the calculator with your real volumes and re-check after a week; for an audit trail, the price API stamps every snapshot with updated_at.
Billing runs per call: the estimated cost is pre-deducted when the request starts, the actual usage is settled when it completes, and any difference returns to your balance. A request that does not complete successfully — upstream error, timeout, no usable output — is refunded automatically. Only successful calls are billed. That policy is what makes aggressive lane competition safe for you as a buyer: a lane that wins your traffic on price has to actually deliver the tokens, or it earns nothing from the attempt.
The marketplace itself charges nothing: no listing fee, no matchmaking commission, no transaction cut, no routing fee. Combined, the three mechanisms — bidding, failure-free billing, merchant-set prices — mean your entire bill is the model cost at the market floor. There is no second line item hiding behind the quote.
See it live: every number on this page is re-derivable from the free public API (GET https://321ai.cc/api/public/model-prices) and the per-supplier quotes on the live marketplace. If a claim on an aggregation platform cannot be re-derived from data you can fetch yourself, treat it as marketing. Tool setup lives in the setup guides, and the FAQ answers the short questions.
Not always. Routing scores price together with live health and task profile, and the stable token group prioritizes stability over price. The default group is the price-first lane; you choose the trade-off per key.
Yes — it is a live market. Conditional routes re-pick the best lane within 5 minutes as prices move, so your traffic follows the market instead of being locked into a stale rate.
No. Failed requests are never billed and pre-deducted quota is refunded automatically. Only successful calls settle at the quoted rate.
On the marketplace page for each model: per-supplier price, upstream self-check records. The catalog floor and supplier counts are also exposed as JSON via the public price API.
Live prices: /en/models/ · billing: /en/pricing · setup: /en/guides/ · FAQ