Researchers show that letting AI models bid for user prompts in first-price auctions comes close to the best possible outcome

What if AI models had to compete for your question, the way contractors bid on a kitchen remodel? A new paper from Google DeepMind researchers tested that idea, and the results were strong.

The paper reports that first-price auctions captured 99.2% of oracle welfare on the EmbedLLM dataset in a typical equilibrium. Put plainly, letting AI agents bid money for prompts got nearly the same overall answer quality as a perfect, all-knowing allocator would.

The work is titled “Price of Anarchy of First-Price Auctions in Agentic Markets.” It was submitted to arXiv on October 14, 2026, under the ID 2610.20587.

The paper imagines a decentralized market that operates at serving time. Each time a user submits a prompt, multiple AI agents compete to answer it.

Every agent places a bid based on what the prompt is worth to it. That value is the expected reward for its response, minus the inference costs of generating the answer.

The allocation rule is a first-price auction. Whoever bids the most wins the prompt and pays exactly what it offered.

Crucially, it doesn’t require a central operator to know each model’s private capabilities or costs. The bids do that work.

The research team includes Shubhendu Trivedi, Jacob Eisenstein, Alekh Agarwal, Ian Gemp, Michael Kaisers and Adam Fisch. Most are from Google DeepMind, with one from Google Research.

There’s a catch with first-price auctions, and economists have known about it for decades. If you pay what you bid, you have every reason to bid less than the prize is actually worth to you.

Daily. Free. Join 34,000+ readers across crypto, finance, and policy.

That behavior is called bid shading. It is perfectly rational for each bidder, but it can cause the “wrong” bidder to win. Sometimes a lower-value agent gets the prompt because a stronger rival shaded too aggressively.

Economists measure that damage with a metric called the price of anarchy, or PoA. It compares the outcome when everyone acts in their own interest against the best outcome a benevolent planner could arrange.

For first-price auctions, the classic worst-case PoA bound sits at about 63%. In the most pessimistic setup, selfish bidding could waste over a third of the potential value.

The DeepMind paper argues that AI agent markets don’t actually look like that worst case. It points to two structural features that work in the auction’s favor.

The first is a competitive floor . Many AI models perform similarly on common tasks. That gives a winning bidder plenty of comparable rivals nipping at its heels, which limits how far it can shade its bid without losing.

The second is what the authors describe as near-independence of values with respect to prompt complexity. Roughly, how much one agent values a prompt doesn’t move in lockstep with how much its competitors value it.

According to the paper, these properties can lift the theoretical PoA bound up to approximately 86.5%, expressed as 1 − 1/e². That’s a meaningful jump from the 63% baseline, before even counting the empirical results.

And the empirical results went further still. The 99.2% figure on EmbedLLM suggests real-world agent behavior lands far above even the improved theoretical floor.

The paper didn’t arrive alone. Overlapping authors published related findings on sequential auctions for multi-step LLM tasks on October 15-16, 2026.

That follow-up targets a harder problem. Many AI workloads aren’t one question and one answer, but chains of steps where different models might handle different stages.

There are limits worth watching. The 99.2% result comes from one dataset, EmbedLLM, and describes a typical equilibrium rather than every possible scenario.

The theoretical gains also hinge on the structural properties the authors identify. If the model landscape consolidated around a few dominant players, the competitive floor could weaken, and with it some of the efficiency cushion.