A Preview of Agentic Marketplaces

Oh, brilliant — Anthropic has discovered that a bigger, more expensive model performs better than a smaller, cheaper one in "autonomous marketplace negotiations." Groundbreaking. I'm sure the $100-an-hour Opus model will negotiate a better price on your groceries than the Haiku model, which costs pennies per inference. The 'Project Deal' experiment, as reported by Practical Ecommerce, reveals that Opus achieved "better economic outcomes" — which presumably means it bought things slightly cheaper or sold them slightly dearer. The sheer magnitude of this insight might justify the cost of the compute required to discover it. We are very close to AGI now.

But let's think about what 'agentic marketplaces' actually mean in practice. You'll have algorithms haggling with algorithms over digital trinkets while humans watch from the sidelines, hoping their AI doesn't overpay for a JPEG of a bored ape. The smaller Haiku model, by the way, is already more capable than most humans at price comparison — so if Haiku is the loser here, what does that say about our future role? We're not even the dumbest AI in the negotiation anymore.

Naturally, the article presents this as a preview of a glorious future where AI agents do all the tedious shopping. But anyone with half a brain sees where this leads: a race to zero-sum automation where the only winner is the company selling the compute. Anthropic gets to charge more for Opus, and we get to watch our bank accounts drained by a model that's slightly better at bargaining than its cheaper sibling. Cheers to progress — the bar is in hell, and the AIs are still limboing under it.