Start hereLive V1

Why a market, not a price list

Inference cost varies wildly across providers. A live auction finds the real clearing price; a posted price list cannot.

The marginal cost of serving one more request is wildly different from one provider to the next. Hardware generation, batching efficiency, GPU utilization at this minute, prefix-cache hit rates, and local energy prices all move it. Yet almost every LLM router charges from a static price list: one number per model, updated whenever someone remembers.

A price list has to be set high enough to be safe for the seller at their worst hour. That gap between posted price and actual marginal cost is pure overcharge, and it lands on you. A live market removes the gap: providers re-quote every couple of seconds as their machines fill and drain, and each request clears at the price competition actually supports at that moment.

Why an auction, specifically

Competition only lowers prices if quoting low is safe for the seller. In a first-price auction it is not: bidders shade their bids above cost and spend their effort guessing rivals. Omnious clears second-score instead. The winner is paid the price that would have tied the runner-up, so bidding below your true cost only wins you losses and bidding above it only loses you profitable fills. Truthful quoting is the dominant strategy, and the market stays honest without policing. The mechanics live in Second-score clearing.

What the target looks like

The design goal is best-execution economics: 20 to 40 percent savings against posted-price routing for equivalent quality, achieved by routing to whoever is genuinely cheapest right now rather than to whoever markets best. The savings are not a promise; they are measured. Every receipt includes the full counterfactual (what each rival quote would have charged for the same request), and your running savings versus posted prices surface as the Price Advantage Index in Receipts & analytics.

the honest caveatA market needs liquidity to beat a price list. V1 launches with allowlisted seed providers, and the auction falls back to a reserve price when fewer than three bidders are live, so an empty book cannot be exploited. The full liquidity plan, including the first-party provider that seeds the book, is in hyper: the first-party provider.

As a customerYou do not have to time the market or pick providers. Send the request; the router finds the cheapest machine that meets your constraints at that instant.

As a providerEfficiency finally pays. If your serving stack clears tokens cheaper than the field, you win fills a price list would never have routed to you.