Giving AI more information may make it even worse at making money.
A Stanford experiment tested 10 frontier large language models across trading, bargaining, and bilateral auctions. The researchers changed just one condition: how much the models knew about their opponents’ true valuations.
The results were quite counterintuitive: with more complete information, some models were actually more likely to lose out on gains.
In the auction experiment, when both buyers and sellers knew each other’s exact valuations, Claude Opus 5.5’s share of potential gains fell from 94% to 64%.
In the trading experiment, different models even showed opposite trends:
GPT-6 Sol’s gains fell from 98% of the maximum to 91%, while Claude Sonnet 5’s rose from 72% to 86%.
In auction scenarios, more information also led models to ask for higher prices. Grok 4.7’s share of gains fell from 97% to 62%.
The same information produced completely different results across models.
This study offers an important reminder for real-world AI agent applications: it’s not enough to test a model’s trading ability alone. You also need to test how it makes decisions under different information conditions.
Especially when you’re asking AI to negotiate, make offers, or manage money on your behalf, more information doesn’t necessarily mean better decisions. Some models may become more aggressive when given more information, ultimately hurting their own returns.
A Stanford experiment tested 10 frontier large language models across trading, bargaining, and bilateral auctions. The researchers changed just one condition: how much the models knew about their opponents’ true valuations.
The results were quite counterintuitive: with more complete information, some models were actually more likely to lose out on gains.
In the auction experiment, when both buyers and sellers knew each other’s exact valuations, Claude Opus 5.5’s share of potential gains fell from 94% to 64%.
In the trading experiment, different models even showed opposite trends:
GPT-6 Sol’s gains fell from 98% of the maximum to 91%, while Claude Sonnet 5’s rose from 72% to 86%.
In auction scenarios, more information also led models to ask for higher prices. Grok 4.7’s share of gains fell from 97% to 62%.
The same information produced completely different results across models.
This study offers an important reminder for real-world AI agent applications: it’s not enough to test a model’s trading ability alone. You also need to test how it makes decisions under different information conditions.
Especially when you’re asking AI to negotiate, make offers, or manage money on your behalf, more information doesn’t necessarily mean better decisions. Some models may become more aggressive when given more information, ultimately hurting their own returns.