First-Price AI Auctions Deliver 86.5% Optimal Efficiency When Models Compete

A new paper analyzes an AI serving market in which agents bid for the right to answer each prompt. The highest bidder wins, pays its bid, and produces the response. The authors ask how closely this simple mechanism approaches the most efficient allocation when agents bid strategically.

The paper proves that first-price auctions can achieve substantially better welfare in agentic markets than the classical worst-case guarantee suggests. Under market structures the authors associate with AI serving, the guaranteed share of optimal welfare rises from about 63.2% to 86.5%. Stronger competition can push the bound toward 100%.

When a prompt becomes a lot

An agentic market turns each incoming prompt into a serving opportunity. Agents estimate their value for handling it, accounting for expected reward, answer quality, inference cost, and any other utility represented by the market. A coding specialist may value a debugging request highly, while an inexpensive general model may have the strongest net value for a simple factual query.

A first-price auction can allocate the request in one round:

1. Eligible agents observe the prompt or a representation of it.
2. Each agent submits a private bid.
3. The highest bidder receives the request.
4. The winner pays its bid and serves the response.

The auction does not require the routing service to know every model’s private capabilities or costs. It still requires infrastructure to collect bids, select a winner, settle payment, and enforce delivery.

The paper defines social welfare as the value created by the allocation. Auction payments count as transfers within the market, so efficiency depends on whether the request reaches the agent with the highest value. Revenue, latency, answer quality, and operating cost can affect valuations, but they are separate evaluation dimensions unless the market incorporates them explicitly.