# Google DeepMind and Google Research Find AI Auctions Reach 99.2% of Oracle Welfare

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/38896
Published: 2026-10-23T10:51:31.000Z
Updated: 2026-10-23T10:51:31.000Z
Section: Technology

> In simulations on more than 30,000 prompts, 110 models reached 99.2% of an oracle allocation’s modeled welfare, measured using estimated correctness and a model-size cost proxy.

Researchers at Google DeepMind and Google Research found simulated AI prompt auctions preserved 99.2% of an oracle allocation’s modeled welfare, suggesting a possible way to route requests among competing models.

The model used first-price auctions for each prompt: Models submitted bids, and the highest bidder received the request and paid its own bid. At a typical approximate equilibrium, 110 bidders reached 99.2% of the welfare produced by an oracle assigning each prompt to the model with the highest estimated value. A second simulation reached 98.8% in an approximate adversarial equilibrium. In that scenario, a bidder could choose a response that lowered welfare only if its private payoff stayed within 5% of its best possible payoff.

Welfare was a modeled measure, not an answer-quality score. Researchers estimated each model’s chance of answering correctly from its 16 nearest neighbors in Gemini’s prompt-embedding space, then subtracted a cost proxy scaled to its parameter count. The formula assigned one cost unit per trillion parameters; it did not measure inference costs in dollars.

The dataset contained more than 30,000 prompts across math, programming, question answering and reasoning, with scored responses from 112 models. Two models with unknown parameter counts were excluded from the auction. Choosing a model separately for each prompt cluster delivered 89.4% of oracle welfare; using the model with the best overall average delivered 78.1%.

Separate theoretical bounds put first-price auctions at about 63.2% of optimal welfare in the general worst case. Under a conditional-independence assumption, the guarantee approaches 86.5%. For the full dataset, the competitive-floor bound was 79.9%, with a 67.6% minimum for the weakest prompt cluster. The paper’s authors wrote that they “do not expect it to be tight.”

The results offer a mechanism-design idea for projects exploring ways to coordinate AI agents, including in crypto ecosystems. The simulations did not use a blockchain, token incentives or crypto payments, and they do not establish how a deployed marketplace would perform.