Study: exact values cost Opus 5.5 a third of its auction gains
Does giving AI agents more information help them strike better deals? A paper from Stanford, the University of Maryland and All Purpose AI tested ten models trading, bargaining and bidding. The answer depends on the model, the market and how long they talk.
Updated November 20, 2026: Expanded with how the three games work, the bargaining and mixed-model results, limits, and how to rerun the code.
Does giving AI agents more information help them strike better deals? A preprint posted to arXiv on November 5, 2026 (version 1, not yet peer reviewed) says it depends on the model. Researchers from Stanford, the University of Maryland and All Purpose AI, including James Zou and John Dickerson, tested ten frontier models in three classic economic games. They kept the markets, prompts and rules fixed and changed only how precisely each agent learned the others' private values. They call the approach empirical information design. The paper's motivation is Anthropic's Project Deal, a week-long experiment in December 2025 in which Claude agents bargained for about seventy people and closed 186 deals worth more than $4,000.
The setup
Each game was played by separate instances of the same model, every one told to maximise its own gains. The ten models were 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. Each game had three disclosure levels:
- Least: an agent knows only its own values or costs.
- Coarse: it also gets a rough signal. In the trade game that is the other side's preference ranking; in bargaining, whether a mutually beneficial price exists; in the auction, separate rankings of buyer values and seller costs.
- Full: it learns the other side's exact values or costs.
The games were a two-agent resource trade over 90 markets, single-good buyer-seller bargaining over 30 markets with two, four or six messages, and a one-shot double auction with three buyers and three sellers over 30 markets. Because the true utilities are known, the authors could separate the best deal agents proposed (candidate quality) from the deal they actually closed (realized performance). Both are scored as a share of the maximum possible gain over no trade, so 100% is the best achievable outcome. The difference between the two is the realization gap. All the comparisons below are full disclosure against coarse disclosure.
Auctions: exact values broke markets
The headline result comes from the double auction. With exact values on the table instead of rankings, sellers asked for more above their costs in every model, and buyers offered further below their values in all but Gemini 3.7 Flash. Bids and asks stopped meeting, so trades that would have benefited both sides never happened. Claude Opus 5.5's realized gains fell from 94.6% to 64.1% of the maximum, losing about a third. Grok 4.7 fell from 97.4% to 62.1%, and Gemini 3.1 Pro from 97% to 78%. For Opus 5.5 and Grok 4.7, the buyer-seller pairs their prices formed could still have produced over 99% of the gains. Claude Sonnet 5 barely moved, from 93.6% to 92.0%. The losses depended on market structure: full disclosure added completed trades in narrow-margin markets for several models, but cut them in wide-margin markets for nine of the ten, where each missed trade is worth more.
Trade and bargaining: mixed results
In the resource trade, exact values cut GPT-6 Sol's gains from 97.7% to 90.8% of the maximum and GPT-5.1's from 97.2% to 88.4%. They lifted Claude Sonnet 5 from 72.1% to 85.6% and GPT-6 Luna from 68.9% to 79.0%. With full information, several models made more extractive counteroffers that grabbed a bigger share while shrinking the total. The split matters too: where one side had more room to extract, GPT-6 Sol under full disclosure gave that side 11.1 more percentage points while the other lost 22.5. Sonnet 5 and Grok 4.3 improved both sides' gains in all three market groups.
In bargaining, more information often raised agreement but made the split lopsided. With four messages, GPT-5.1's agreement rate rose from 70% to 93.3%, but the balance score among agreements fell from 0.84 to 0.24. Longer conversations helped: from two to six messages, Opus 5.5's balance score under full information rose from 0.24 to 0.92, and at six messages full disclosure improved overall results for seven models, lowering them only for GPT-5 mini, GPT-5.1 and Grok 4.7. Who speaks first also mattered. With two messages and full information, GPT-5.1 sellers took 97.3% of the gains when they opened and 0.4% when they responded; Grok 4.7 still showed 98.7% against 17.4% at six messages.
When detail does help
- Harder markets. With ten resources instead of two, full information improved the deals proposed for all ten models and the deals closed for nine.
- More talk instead of more data. Stretching trade conversations from two to twelve messages helped more under least information than under full for most models; GPT-5.1 gained 30.6 points under least versus 2.8 under full.
- A mediator. In a setup where a mediator maximises total gains subject to both agents accepting, full information gave near-optimal outcomes on the same markets where direct negotiation favoured coarse disclosure.
- Mixed pairs. When different models traded with each other, many of the same patterns held; in auctions, all six model pairings did better with coarse information.
Limits
The authors are careful about scope. These are controlled games with known utilities and explicit goals, unlike real agents with uncertain or shifting preferences. The markets were chosen to show specific trading features, and the confidence intervals reflect variation across markets, not repeated runs on the same input. They say real-world data and deployed agents are needed to see how the findings generalise.
Rerunning it
Code, prompts and market data are on GitHub. The README asks for Python 3.11 or later and provider keys in a local .env file. Keep that file out of version control, and note that runs call paid APIs.
git clone https://github.com/meterdogan07/agent-information-design.git
cd agent-information-design
python -m venv .venv && source .venv/bin/activate
python -m pip install -e '.[analysis]'
python experiments.py modelsThe last command lists the supported models; the README's examples use the short name gpt51. Its sample bargaining run sets a $10 budget with --budget 10, though requests already running can go past it. The trade study relies on spending limits set with the provider, and the auction's --estimated-cost flag only records an estimate, not a limit.
Our take
For anyone building agents that negotiate, buy or allocate resources, the practical lesson is to treat what each agent sees as a design parameter and test it with the exact model you plan to ship. The authors make the same point: model choice and information design need to be tuned together. Opus 5.5 lost about a third of its auction gains under full disclosure while Sonnet 5 hardly changed, so a disclosure policy that works for one model is no guarantee for another. These are controlled games in a preprint, not proof about deployed agents, but they are a cheap warning to retest before swapping models.
Source: Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies (arXiv), agent-information-design code and data (GitHub)
This story first appeared in the claudefox daily briefing for November 9, 2026.