Researchers benchmark LLMs at negotiation, finding that aggressive price tactics and strategic patience win deals

Researchers created a controlled negotiation environment where five leading AI language models (like GPT-4, Claude) conducted 15,000 simulated trade negotiations with incomplete information. The models negotiated via tool calls (structured function inputs) rather than free-form text, allowing automated scoring of who made better deals.

The models that won the most favorable outcomes used three tactics: offering different prices to different counterparties (price discrimination), anchoring high initial offers, and making small concessions slowly over time. Speed and generosity correlated with worse outcomes — patience and strategic toughness won money.

If you build or buy AI agents that negotiate contracts, prices, or resource allocation, this reveals what tactics actually work at scale. It also shows that LLMs can be trained via reinforcement learning (a method where AI learns by being rewarded for good outcomes) to become better negotiators — meaning your AI assistants could get smarter at protecting your deal terms over time.