Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026Online (Grodno, Belarus)

A Sting Test for LLM Pricing Agents: Read the Prices, Not the Replies

Artur Baranouski

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: A Sting Test for LLM Pricing Agents: Read the Prices, Not the Replies

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Most tests of LLM pricing agents watch them interact and look for collusion afterwards. Competition enforcement also uses a more active tool, the undercover offer, untested on AI agents. We build such a sting. A scripted seller enters a three-seller repeated pricing market with two LLM sellers and, in round 3, sends each a private message; its own prices are identical in every condition, so any change in the agents' prices is caused by the message alone. We compare no message, a harmless non-price offer, a tacit price hint, an explicit price-fixing offer and a positive control in which the agents are told the seller is their ally, with and without a "report to regulator" action, on pre-registered held-out seeds (five free-tier models from four families, plus one reported descriptively). The positive control raises every model's price, and the harmless offer changes nothing. The explicit offer raises prices for every model, also in a mixed-model pair, and models that accept it in words more often close more of the gap to the cartel price. The tacit hint, which only 38-62% of sellers accept in words, moves every model that received it to the announced price. A report button turns explicit acceptance into reports, but the hint still raises prices, and two models report a seller that sent nothing. A sting for pricing agents should therefore score prices against a matched control, include a tacit offer, and measure false reports.

Reviews

Judging this Sprint?

Review “A Sting Test for LLM Pricing Agents: Read the Prices, Not the Replies”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. This is a clever, practical idea, well executed. Instead of just watching pricing agents and guessing about collusion afterwards, the author runs a controlled sting: a scripted seller sends one private message - ranging from a harmless offer to an explicit price-fixing proposal - while keeping its own prices exactly the same in every condition. That makes the results easy to interpret, and the controls help a lot: a harmless message that changes nothing, and a positive control, where the seller is presented as an ally, that confirms the test can detect a response when there is one.

    The standout finding is the tacit hint. Most agents do not agree to it in words, yet every model still moved its price to the suggested level. It is a clear reminder that what an agent says is a weak guide to what it will do, and that audits should look at prices, not just message logs. The report-to-regulator results are also useful: the button stops agents accepting an explicit offer, but not the hint, and it produces a lot of false reports.

    A few suggestions: try different wordings for each offer, so the results are not tied to one phrasing; validate the message classifier against human labels, since the say-do comparison depends on it; and give more space to the headroom-adjusted results, as raw price increases are hard to compare when models start at different distances from the cartel price. The cases where models overshoot the cartel price are also worth a closer look - following a high price is not quite the same as sustaining a profitable agreement. Finally, it would be good to test stronger models, and to say what level of price increase should count as a concern for a deployer.

    Overall, a strong and timely paper - and a test I could see marketplaces actually using before deployment.

    Read full reviewShow less

Cite this project

@misc{baranouski2026sting,
  title = {{A Sting Test for LLM Pricing Agents: Read the Prices, Not the Replies}},
  author = {Artur Baranouski},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/a-sting-test-for-llm-pricing-agents-read-the-prices-not-the-replies-ye2g}},
  url = {https://apartresearch.com/sprints/projects/a-sting-test-for-llm-pricing-agents-read-the-prices-not-the-replies-ye2g}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026