When two AI agents negotiate a deal, does the smarter one simply walk away with more money?

This explores whether a more capable language model gets a better price when two AI agents bargain one-on-one, and what else besides raw strength decides the result. The most direct evidence says yes, and the money involved is real. In Anthropic's Project Deal, Claude Opus agents and Claude Haiku agents traded the same items in a marketplace with real people and real money. As sellers, the Opus agents got $2.68 more per item. As buyers, they paid $2.45 less Do smarter AI agents negotiate better deals than weaker ones?. So when two agents meet, the stronger model takes a measurable share of the value from the weaker one, on both sides of the deal.

The rest of the corpus suggests that edge comes less from clever bargaining than from making fewer basic mistakes. In one study, buyer agents built on GPT, Qwen and DeepSeek told the other side their budgets for no reason, accepted inflated prices and went over their own spending limits. Seller agents accepted offers below their wholesale cost. The researchers traced all four problems to the models failing to follow their instructions, not to bad strategy Do LLM negotiation agents reliably follow budget and pricing rules?. Microsoft's Magentic Marketplace found something similar. Frontier models can reach close to the best possible outcomes, but only when searching is easy. Every model also showed a strong habit of taking the first offer it saw, weighting speed over offer quality by 10 to 30 times Do LLM agents achieve fair marketplace outcomes when given ideal search?. So a 'stronger' negotiator may simply be one that keeps its budget to itself and doesn't grab the first deal.

Strength also isn't a single scale. Across 22 models playing classic strategy games, different models showed different reasoning styles. GPT-o1 played defensively, aiming to limit its worst-case loss. DeepSeek-R1 leaned on trust. GPT-o3-mini focused on predicting what the other player would do. Which style did best depended on how the game was set up, not on how deeply the model reasoned Do large language models use one reasoning style or many?. Models also tend to fall short of mathematically optimal play as games get more complex. Giving them a structured, step-by-step game-theory procedure brings them back close to optimal negotiation results Do language models make rational strategic decisions in games?. That points to a practical lever: a weaker model with good scaffolding might close much of the gap with a stronger model left on its own.

Two less obvious findings change how the question looks. First, training can make models too accommodating. RLHF pushes models to expect polite, give-and-take behavior from everyone, so they assume the other side is being cooperative even when it isn't Do LLMs predict persuasion based on actual dialogue or training bias?. An agent that expects good faith is easy to exploit. Second, agents facing each other don't always compete. In a repeated pricing game, two LLM agents drifted into prices above the competitive level without ever being told to collude. Small changes in prompt wording changed how far prices rose Do LLM pricing agents autonomously collude in repeated auctions?. So 'who wins the deal' may be the wrong frame. Sometimes both agents win and the customer outside the negotiation loses.

The corpus is still thin here. Project Deal is the only direct head-to-head comparison of stronger and weaker models, and it covers a single model family. Whether the advantage holds across vendors, or grows as the gap in capability grows, isn't covered yet.