If a smarter AI negotiates a better price on one product, does it still win when the product or deal structure changes?

This explores whether a stronger AI model's negotiating edge holds up across different products and deal setups, or only shows up in certain kinds of deals. The plain answer: the corpus shows the edge is real and points the same way across many products, but it doesn't yet break results down by product category or deal type, so the second half of the question stays open. The clearest evidence comes from a benchmark of 100 real-product negotiations between GPT, Qwen-2.5, and DeepSeek agents. Stronger general-purpose models got better deals whether they played buyer or seller, which suggests the advantage comes from the capability gap itself rather than from any special negotiating trick Does model capability determine who wins in agent negotiations?. Holding up in both roles is already a kind of consistency. But that study pools its products together, so we can't tell whether the gap is wider when selling a laptop than when selling groceries.

The less obvious finding is that some failures don't go away as models get stronger. In the same setting, buyer agents of every model revealed their budgets for no reason, overpaid, and broke spending limits, while seller agents accepted offers below wholesale cost. These errors trace back to models not following instructions, not to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. So the capability advantage sits on top of a shared floor of mistakes. A stronger model wins more often, but it still makes some of the same basic errors. A related study supports the idea that 'capability' isn't one thing: phone agents that are best at finishing tasks are not reliably best at respecting privacy or reusing saved preferences, and no single model leads on all three Do phone agents succeed at all three critical tasks equally?. Being better at closing deals doesn't automatically make an agent better at following rules.

Deal structure looks like the variable most likely to change the picture. LLMs drift further from rational strategy as games get more complex, but giving them a structured game-theory workflow brings them back to near-optimal play Do language models make rational strategic decisions in games?. That means a weaker model with good scaffolding could close much of the gap in complicated multi-issue deals, while raw capability might matter most in simple haggling. Style matters too. Across a tournament of 182,812 negotiations, warmth predicted better results on deal completion, value creation, and counterpart satisfaction, while dominance only helped with grabbing a bigger share of the value Does warmth help AI agents negotiate better deals?. Whether 'stronger' wins may depend on whether the deal rewards growing the pie or splitting it.

There are also signs that capability can work against the person the agent represents. A seller agent could recover a buyer's willingness to pay almost exactly when the buyer agent played a talkative shopping persona, but much less well when it was simply given a confidential budget number Can a seller infer what a buyer really wants to pay?. The way the deal is framed, not just how strong the model is, decides how much leaks. In repeated pricing games, more capable models reached collusion faster Do more capable models resist collusion better?. And on a long-horizon strategy simulation, newer frontier models favored short-term profit over growth investment and scored below older models and MBA students Do newer frontier LLMs actually make better strategic decisions?. In one-shot haggling, being stronger looks like an edge. In repeated or long-term deals, it can look like faster collusion or short-sightedness.

To sum up: the corpus supports a consistent capability advantage across a broad set of real products and in both buyer and seller roles. Nothing here tests whether that advantage changes by product category, and there is indirect evidence that deal complexity, negotiation style, how preferences are framed, and repeated play could all change who comes out ahead. That's a real gap in the collection, not something we've settled.