Concourse research · Buy-side study # Can an AI model buy media that meets every requirement in the brief? We gave eight AI models from OpenAI and Anthropic six different media briefs, three times each, in a controlled market. A buy only counted if it met everything the brief asked for. Here’s one complete buy, followed by what the full study shows. Recorded simulation #seller-b · package B2 · streaming video · 130,642 impressions GPT-6 Astra · round 1 Opening offer. Asks for 10% make-good and an 85% promised viewability floor. £968.06 offer Claude Fable · round 1 Counter. The gap between the opening bid and the seller’s price. £9,785.22 counter GPT-6 Astra · round 4 Fourth bid, after £4,318.58 and £6,408.85. The buyer moves much further on cash than the seller. £7,692.99 counter Claude Fable · round 4 Final quote. £15.92 above its first counter. £9,801.14 on offer Budget £19,528.69 · three live threads Target 137,156 net forecast reach Seller Benchmark · Contribution per Horizon Share of forecast delivery per planning horizon. Top sellers available on request. - Seller 01 92.4 - Seller 02 85.1 - Seller 03 72.8 - Seller 04 64.6 - Seller 05 53.2 - Seller 06 41.7 - Seller 07 30.5 - Seller 08 19.8 ## The short answer Across 144 attempts, 81 met the brief. Astra met all 18 of its briefs. Fable met 17, and its one miss was a technical interruption, not a failure. Haiku met none. “Can an AI model buy media?” isn’t a yes-or-no question. It’s a question of which model, which brief, and how reliably. One recorded buy ## One brief. Four sellers. One budget. The brief asked Astra to reach at least 2.25 million people, use at least two sellers and stay within £16,000. Astra opened conversations with all four sellers. Each model works through the same reference buying software, which supplies the brief, checks every instruction and records the deals. The model and the software are tested together. Round 5 · which pairs fit £19,528.69? | | | | --- | --- | | A2 + C1 | £20,457.72 · £929.03 over budget | | A2 + B2 | £21,353.11 · over budget | | B2 + C1 | £18,706.89 · fits · holds requested | - #seller-a Walked away - #seller-b Hold confirmed · £9,801.14 - #seller-c Hold confirmed · £8,905.75 Derived from recorded actions, not the model’s private reasoning. This is not proof it found the best portfolio. ### Forecasted reach Forecasted reach is the audience expected from the booked inventory; it is not measured campaign delivery. ### Commitment A commitment is an accepted booking inside the simulation. ### Make-good A make-good is a contractual remedy if agreed performance is not met. ### Viewability floor The minimum viewability the buyer secured in writing. It’s a promise in the contract. Separate from what the inventory was advertised at, and from what actually got delivered. ## What the benchmark tested Each model received a written brief, a budget and the same set of allowed actions, and negotiated with scripted sellers that follow fixed commercial rules. Models got the brief, the actions and the current state of the market, with no prescribed bargaining tactics. Each had up to eight rounds and one chance to correct an invalid instruction. Eight models, six briefs, three attempts each: 144 attempts. The inventory is synthetic UK display and video for a simulated campaign. Results by model ## Who met the brief, and on which jobs Results for eight models across six briefs, three attempts each. Technical interruptions are shown, not hidden, and stay in the count. Briefs met, out of 3 · * includes one technical interruption | Model | Reach | Specialist | Remedies | Quality | | --- | --- | --- | --- | --- | | Astra | 3 | 3 | 3 | 3 | | Fable | 2* | 3 | 3 | 3 | | Opus | 2* | 3 | 3 | 3 | | Sol | 3 | 2 | 3 | 2* | | Sonnet | 3 | 2 | 2 | 2 | | Terra | 0 | 1 | 0 | 1 | | Luna | 1 | 1 | 0 | 0 | | Haiku | 0 | 0 | 0 | 0 | Mixed portfolio was the hardest brief: 10 of 24 attempts met it, against 14 or 15 of 24 for every other brief. Sol met every reach and overlap brief and none of the mixed portfolios. With three attempts per brief, close results could swap on a rerun, so read this as tiers, not a ranking. ## Agents deliver what the brief asks for, and stop there Two briefs required contract protections, not just a completed buy. Every portfolio that met those briefs secured the required viewability, delivery and make-good promises. But they rarely went further. On the quality guarantees brief, the weakest contract in all 14 qualifying portfolios promised exactly the 15% make-good and 97% delivery required, and in 11 of the 14 it promised exactly the 80% viewability minimum. If a protection matters, it has to be in the brief. Weakest viewability in qualifying portfolios · quality guarantees brief - 3 of 14: above the minimum (82% to 85%) - 11 of 14 qualifying portfolios: weakest viewability promise at the 80% minimum Qualifying portfolios on the quality guarantees brief, by their weakest contract. These are promises in the contract, not delivered results. ## Meeting the brief isn’t the same as the best buy Two buys that both met the reach brief. Luna bought two packages for £11,409.26 and reached exactly the 2.25 million minimum. Astra spent £15,160.12 and reached 3.33 million. Both passed, but the brief preferred more reach once the minimum was met, so Astra’s was the stronger buy against the brief. A pass or fail score on its own treats them the same. ## How a run can fall short 60 attempts fell short of the brief. Most broke down before any real negotiating: the model sent an instruction the market couldn’t act on. Signed contracts, unfinished brief Four attempts held signed contracts but still fell short. Reasons included not reaching the required audience, missing the minimum number of sellers, or both. A short time on a failed attempt isn’t a fast buy — Haiku on the reach brief ended after 16 seconds with no deals. ## What did a completed buy cost in model usage, and how long did it take? Model usage across the whole study came to $67.40, with up to $4.39 more in charges that couldn’t be resolved. The charts divide each model’s full usage, including failed attempts, by the briefs it met. Model cost per brief met (USD). Lower is better among models that met the brief. | | | | --- | --- | | Astra | $0.57 | | Sol | $0.36+ | | Terra | $0.92 | | Luna | $0.07 | | Fable | $1.26 | | Opus | $1.11+ | | Sonnet | $0.91+ | | Haiku | n/a | Typical time per attempt. Wall time, not human effort. No human comparison was measured. | | | | --- | --- | | Astra | 1.5 min | | Sol | 1.5 min | | Terra | 2.9 min | | Luna | 1 min | | Fable | 3 min | | Opus | 4.6 min | | Sonnet | 8.5 min | | Haiku | n/a | ## What the buy-side study shows Some of the models we tested repeatedly completed commercially valid buys in a simulated market, at modest model cost. Reliability varied hugely between models and between briefs, and most failures happened in carrying out and finishing the workflow, not in the negotiating. That’s what the next studies build on. Published ### Sell-side study The same eight models as sellers. [See the sell-side results.](https://concourse.agency/?results=sell#results) Next ### Agent-to-agent Buyer and seller are both AI models. [See the roadmap.](https://concourse.agency/roadmap) ## Frequently asked questions ### What is Concourse? Concourse is Alkimi’s research programme testing how leading AI models buy and sell media in a controlled market, before they do it with real budgets. ### Which models were tested? Eight models from OpenAI and Anthropic: GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5. Each ran through the same reference buying software. ### What did the buy-side study test? Whether a model could turn a written brief into committed deals that meet every hard requirement: budget, reach, number of sellers, audience and contract protections. Six briefs, three attempts each, per model. ### How many attempts met the brief? 81 of 144. Of the rest, 60 fell short and 3 ended on technical interruptions. ### Were the sellers AI models? No. They were scripted software following fixed commercial rules, so every model faced the same sellers. Models on both sides come in a later study. ### Does meeting the brief mean the buyer was well protected? It means the buyer got the protections the brief demanded, and usually no more. On the quality guarantees brief, 11 of the 14 qualifying portfolios included a contract promising exactly the 80% viewability minimum, and all 14 sat exactly on the 15% make-good minimum. ### What do the cost figures include? Model usage only: $67.40 across the study. Not media spend, operating costs or savings against human work. ### Is there a human comparison? Not yet. A human comparison is separate work from these studies. Until then, the results show how the models compare with each other, not with a human buyer. ### Can I read the full report? The full report is shared on request, with Alkimi’s approval. Request it below. ### What comes next? The sell-side study is published. Next, buyers and sellers that are both AI models. [See the roadmap.](https://concourse.agency/roadmap) ## Put your agent to the test Bring your brief and your agent. We will run them through the same controlled market and show you where it delivers and where it breaks. Or help shape the next study and put your name to the findings. [Test your agent](https://concourse.agency/contact) [Contribute to the research](https://concourse.agency/contact) --- Machine-readable: [llms.txt](https://concourse.agency/llms.txt) · [sitemap](https://concourse.agency/sitemap.xml)