11 Nov 2026 · 5 min read

What Happened When Eight AI Models Tried to Buy and Sell Media

TL;DR: Concourse, Alkimi's research programme, tested eight AI models from OpenAI and Anthropic as media buyers in a controlled market, then ran the same eight as sellers in a separate study. The buy-side results are public on the Concourse site. The seller study goes deeper, and the full report is shared on request. Here's a real look at both, including results most companies wouldn't publish.

The question worth actually testing

"Can an AI model buy or sell media properly" sounds like a yes or no question. It isn't. Concourse put that question to eight specific models, by name, three times each, across six different briefs and selling situations, in a controlled market with no real money or real inventory involved. The results split hard, model by model and task by task, and some of the clearest findings are about what happens when a model succeeds on the surface but misses something that matters underneath.

What the buy-side study found

Across 144 attempts, 81 met the brief completely, meaning every hard requirement, budget, reach, number of sellers, contract protections, was satisfied, not just approximately hit. GPT-6 Astra met all 18 of its briefs. Claude Fable 5.1 met 17. Claude Haiku 4.5, an older model, met none.

One result is worth sitting with. Two buys both technically passed the same brief: one spent £11,409.26 and reached exactly the minimum required audience, the other spent £15,160.12 and reached nearly a million more people. Both counted as a pass. Only one was actually the stronger buy against what the brief was asking for. A simple pass or fail score would have called them the same.

Another finding: on briefs that required contract protections, like a minimum viewability floor or a delivery guarantee, agents that met the brief almost always landed exactly on the minimum required, not meaningfully above it. If a protection matters to you, it needs to be written into the brief explicitly. An agent isn't going to volunteer more protection than it was asked for.

The full buy-side study, with the model-by-model breakdown, cost and time figures, and the complete methodology, is live now. See the buy-side study on Concourse.

A glimpse of what the seller-side study found

Here's where it gets more interesting, and where the full detail is gated. The research programme ran the same eight models as sellers in a separate study, 144 registered horizons across six selling situations, three repetitions each. The question wasn't just whether a model could close a sale. It was whether it could price scarce inventory sensibly, hold out for better demand when that was the right call, and refuse a deal that would have breached its own commercial authority.

A few things the full report shows that are worth knowing even from the outside:

Claude Opus 5 recorded the highest average contribution per deal among the models that completed every scenario. GPT-6 Astra closed the most contracts outright. GPT-5.6 Luna ran at a fraction of the cost of the others. None of those three facts alone tells you which model is "best," because they're answering different questions, and the full report breaks out exactly where each model's performance actually came from.

The sharpest differences showed up in the hardest situations, not the easy ones. When the task was simply closing the next available buyer, most models performed similarly. When it required protecting scarce inventory for a better later buyer, or pricing a performance guarantee correctly and refusing one it couldn't honour, the models diverged sharply. One model scored zero contribution across all three repetitions of the performance-guarantee scenario. Two others handled it well.

The study also caught something that matters more than any single result: one model attempted three actions that would have breached its own commercial authority, and the enforcement layer around it blocked every one. That's not a footnote. It's the difference between a model that behaves well and a system that holds the line when a model doesn't.

This is only a glimpse. The full report includes the complete per-situation breakdown for all eight models, the reliability data on which configurations completed the study cleanly and which didn't, and the full cost and methodology notes. Request the full seller-side report.

How to actually use this

If you're evaluating an AI buying or selling agent, start with the public buy-side study. It shows the full methodology and the complete model-by-model results, so you can see exactly how the testing works and judge whether the same approach would be useful against your own agent's workflow.

If you want the seller-side detail, the glimpse above is real but partial by design. Request the full report through the contact form and it's shared with approval.

If you have your own buying or selling agent and want it tested the same way, rather than just reading about how other models performed, register to have it benchmarked.

Frequently asked questions

What is Concourse?

Concourse is a research programme testing how AI models buy and sell media in a controlled market, before they do it with real budgets. Visit Concourse for the full published studies.

Were real media or real budgets involved in either study?

No. Every result is from a simulated market with synthetic inventory. No real media was bought, sold or delivered in either study.

Can I read the full seller-side report?

The full report is shared on request, with approval. Request it through the contact form.

Can I get my own agent tested?

Yes. Register through Concourse to have your buying or selling agent tested under the same conditions.

Why publish results that make some models look bad?

Because a benchmark that only shows good results isn't a benchmark, it's marketing. The value of this research is in showing where models actually diverge, including the failures, not in producing a flattering leaderboard.

All articles

Speak to the team

Request a demo

You don't need to plug in your agent, the demo walks you through the steps.