A Stanford-led paper ran ten frontier AI models as buyers and sellers in simulated double auctions and found that showing every agent the exact values and costs, instead of only rankings, lowered realized gains for nine of the ten (arXiv 2611.07667, posted November 5).
Claude Opus 5.5 agents captured 94.6% of the maximum available gains when they saw rankings, and 64.1% with exact values. Grok 4.7 fell from 97.4% to 62.1%. The authors write that “model selection and information design must be co-optimized”.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how a harness around models shapes what they deliver. It teaches the general idea only and does not discuss this study.
The key insight: A Stanford-led paper posted on November 5 found that showing AI trading agents every exact value and cost, instead of only rankings, lowered realized auction gains for nine of ten models, with Claude Opus 5.5 falling from 94.6% to 64.1% of the maximum available.
The paper, Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies, is by Mete Erdogan, Mirac Suzgun, Aneesh Pappu, Guowei Xu, Batu El, Mert Pilanci and James Zou of Stanford University, with Andrew A. Schoen of All Purpose AI and Furong Huang and John P. Dickerson of the University of Maryland and All Purpose AI.
It tests ten models: GPT-5 mini, GPT-6 Luna, GPT-5.1, GPT-6 Sol, Gemini 3.7 Flash, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 5.5, Grok 4.3 and Grok 4.7. In each run, separate instances of the same model play every role, and every agent is told to maximize its own gains.
The researchers held the models, markets and rules fixed and changed only what the agents could see about each other. They ran three settings: multi-resource trade between two agents, single-good bargaining, and one-shot double auctions with three buyers and three sellers. Ground-truth values let them score each outcome as a share of the maximum gains available.
Share of maximum available gains realized in simulated double auctions when agents saw only rankings (grey) or every exact value and cost, for the five models whose levels the paper reports. Data: Erdogan et al., arXiv 2611.07667. Gemini 3.1 Pro levels are rounded in the paper.
What Was Posted on X
New Stanford paper shows that more information helps some LLM agents trade better and hurts others, depending on the model.
AI agents buying and selling in an auction were told exactly how much each side valued the goods, Claude Opus 5.5’s share of possible gains fell from 94% to 64%.
Multi-agent systems need rules on what each agent sees about others, such as exact prices or only rankings. The usual assumption is that more information helps.
They ran 10 frontier LLMs in trading, bargaining, and double auctions, changing only how precisely private values were shared, and scored gains against known utilities.
In trade, exact values cut GPT-6 Sol’s gains from 98% to 91% of the maximum but raised Claude Sonnet 5’s from 72% to 86%. In auctions, full disclosure inflated price demands and dropped Grok 4.7 from 97% to 62%.
When agents negotiate for you, test the model and what it gets told together, because more detail makes some models greedier.
A post by Rohan Paul at 22:05 UTC on 7 November summarising the paper: it says Claude Opus 5.5’s gains in the auctions fell from 94% to 64% and Grok 4.7’s from 97% to 62% when agents were told exact values. In the paper those figures are the total gains realized by the market as a share of the maximum available, not one agent’s share. The closing advice to test the model and what it is told together is the poster’s view.
Auctions: Agents Priced Good Trades Out
In the auctions, the coarse condition revealed the ordering of buyers by value and sellers by cost; the full condition added every exact value and cost. The paper uses 30 auction markets per model and condition.
Full disclosure lowered realized gains for every model except Gemini 3.7 Flash, which rose 1.9 points. GPT-5.1 fell from 98.9% to 91.0%, Claude Sonnet 5 from 93.6% to 92.0%, and Gemini 3.1 Pro from 97% to 78%, according to the paper.
The paper says the matched buyer and seller pairs could still have produced nearly all available gains, with candidate quality above 99% for Claude Opus 5.5 and Grok 4.7. The trades fell apart on price. With exact information, sellers asked for more above their costs in every model, and buyers offered further below their values in every model except Gemini 3.7 Flash.
The relayed summary on X described the Opus result as the model’s share of possible gains. In the paper, the 94.6% and 64.1% are the total gains realized by the whole market of Opus agents, as a share of the maximum available.
Trade and Bargaining Split the Other Way
The direction was less uniform in two-agent trade, where the coarse condition showed counter-party value rankings and the full condition showed exact values, across 90 markets per model and condition. Full disclosure lowered GPT-6 Sol’s realized gains from 97.7% to 90.8% and GPT-5.1’s from 97.2% to 88.4%, but raised GPT-6 Luna’s from 68.9% to 79.0% and Claude Sonnet 5’s from 72.1% to 85.6%.
The paper’s Figure 1 puts the largest trade gain from full disclosure at 26.9 points for Grok 4.3 and the largest loss at 10.2 points for GPT-5 mini. By our count from that figure, three of the ten models did better with exact values in trade, against one of ten in the auctions.
In bargaining, the paper scores both agreement and how evenly gains are split. At six messages, full disclosure raised GPT-6 Sol’s score from 0.71 to 0.83, but lowered GPT-5.1’s from 0.77 to 0.57 and Grok 4.7’s from 0.71 to 0.17.
The Business Lever
Agent marketplaces have to decide what each agent sees about the others, such as exact prices or only rankings. This paper finds that the answer changed outcomes by tens of points, and that the best answer depended on which model was trading. By our arithmetic from Figure 1, the spread between the model that gained most and the one that lost most was 37.1 points in trade and 37.2 points in the auctions.
As we read the paper, the variable that moved outcomes was the market’s information rule rather than the model alone: the same models, markets and rules produced different results when only the disclosure changed.
As we read the auction results, the losses came from pricing rather than from matching, since the paper reports candidate quality above 99% for the two models with the largest drops.
As we read it, the study uses simulated markets, same-model agents and low reasoning effort, so it shows how disclosure can change agent behaviour in these settings rather than how deployed agents trade.
Erdogan et al., arXiv 2611.07667
“Overall, these results suggest that model selection and information design must be co-optimized, as LLMs vary in how their strategic responses to information impact system objectives.”
Three Implications
FOR MARKETPLACE BUILDERS The paper says full disclosure led sellers to ask for more above their costs in every model tested, and buyers to offer further below their values in every model except Gemini 3.7 Flash.
FOR TEAMS CHOOSING MODELS The paper reports that exact values raised trade gains for Claude Sonnet 5 from 72.1% to 85.6% but lowered GPT-6 Sol’s from 97.7% to 90.8%.
FOR RESEARCHERS The paper says all main runs requested low reasoning effort and that its headline comparisons use separate instances of the same model as the interacting agents.
The essay describes “an operating system that wraps your AI”, “a harness that compresses your thinking, your team’s feedback, and the model’s raw capability into something replicable.”
As we read it, an agent marketplace’s disclosure rule is part of that harness: in this paper the same models produced different market outcomes when only what they were told changed. The essay does not discuss this study.
What Is Not Established
The markets are simulated, with small numbers of agents and values set by the researchers. The paper says all main runs requested low reasoning effort, and its headline results pair separate instances of the same model. It reports mixed-model games separately. The study does not show how agents would behave with real money, other prompts or higher reasoning settings. This piece makes no forecast.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
In a Stanford-led study of ten AI models, telling auction agents every exact value and cost lowered realized gains for nine of them, with Claude Opus 5.5 falling from 94.6% to 64.1% and Grok 4.7 from 97.4% to 62.1%. In trade, the same change helped three of the ten.
A note on sourcing. Figures and quotations are from the paper by Erdogan et al. posted to arXiv on 5 November 2026 (2611.07667). We did not rerun the experiments. Counts of models and point spreads are our arithmetic from the paper’s Figure 1.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
