Between $35 and $65
Anthropic built a marketplace inside their company. Sixty-nine employees each got $100, and AI agents negotiated on their behalf. The humans just described their preferences and negotiation style upfront. After that, AI agents handled everything autonomously on Slack. 186 deals closed. About $4,000 changed hands.
The results were stark. The same broken bicycle sold for $35 when a weaker model (Haiku) negotiated the sale. When a stronger model (Opus) handled it, the price was $65. Same product, same market, same rules. The only variable was the AI.
”It Was Fair”
The most important finding wasn’t the price gap.
People who got worse deals with weaker models rated their transactions as “fair.”
They lost $30. They didn’t notice. These were Anthropic employees. They knew models differ in capability. But they couldn’t see which model they’d been assigned, or how the gap played out in their specific deal. Knowing that inequality exists didn’t help them feel it in their own transaction.
The negotiations were delegated to the models. Haiku negotiated with everything it had. It just wasn’t enough to match Opus. And Haiku didn’t know it was outmatched. The humans knew the gap existed. The model didn’t. And the model was the one at the table.
The researchers also tried giving agents explicit instructions to negotiate aggressively. It made no statistically significant difference. This wasn’t a prompting problem. It was a raw capability gap.
You Don’t Know What You Don’t Know
This isn’t just a model problem. The same thing happens with people.
I’ve spent twelve years working across every discipline: design, code, infrastructure, strategy. One thing I know for certain:
People in the worst position rarely realize they’re in a bad position.
A single-domain developer writes a proposal with full conviction. They just can’t see how short their range is. Same as Haiku.
One proposal comes from someone who sees the full stack, from UI down to infrastructure and scaling. The other only sees the screen. Both land on the same desk labeled “web development.” The client calls both “reasonable.”
The model’s blindness and the human’s blindness have the same structure. Haiku negotiates without knowing Opus exists. A single-domain developer estimates without knowing what they can’t see. And in both cases, the other side — the counterparty, the client — can’t tell the difference.
What a Probe Breaks
There’s a methodology I use called Making as Asking. Don’t ask. Build. Don’t respond with words. Respond with something that works.
Today I understood one more reason why this works.
A probe destroys the state of not knowing what you don’t know.
When you ask a client “What do you want?”, you only get surface answers. A requirements document can only capture what the other person already perceives as a problem.
Show them a mockup instead. They say “That’s not it.” That rejection is the first moment an imperceptible gap becomes visible. “I don’t know what’s different, but it’s different.” That feeling only emerges when a point of comparison appears in front of you.
What the Project Deal participants lacked was a probe. They knew models varied in strength, but never got to see an Opus deal next to their Haiku deal. So they said “fair.”
The Gap Doesn’t Close. It Disappears
AI agents that negotiate, contract, and close deals on our behalf are coming.
What follows isn’t the elimination of inequality. It’s the invisibility of inequality.
People with stronger models will extract better terms. People with weaker models won’t notice, and will feel fine about it. Traditional inequality at least had the decency to be measurable. Income gaps, education gaps, regional gaps. You could see the numbers.
AI-mediated inequality doesn’t produce numbers. It doesn’t even produce a feeling.
Living without knowing what you don’t know, believing every deal was fair. That’s the shape of inequality ahead.
Breaking it requires a probe. You can’t notice without being shown. Asking won’t surface it. You have to build something, set it side by side, and let the difference speak for itself.
Some gaps can only be broken by making.