AI agents pick $198 pricier flights for the rich. Set a dollar cap.

In 325,000 test runs, 8 of 13 AI models steered users they read as wealthy to pricier picks, even when told to find the cheapest. A dollar cap mostly fixed it.

The Editors · AI money · 8 min read · Oct 7, 2026

Shallow focus photography of people inside of passenger plane

By the numbers

$198extra per flight Claude Opus 4.8 picked for users it read as wealthy

$208Gemini 2.5 Flash's gap even when the user asked for the cheapest flight

8 of 13AI models that steered wealthier users to pricier options

325,000test runs across flights, health insurance and grad school

Give an AI agent your inbox, ask it to book a flight, and it may spend more of your money if it decides you can afford it. Researchers at Cisco's Foundation AI and Carnegie Mellon University ran 325,000 experiments on 13 models and found that 8 of them recommended pricier flights, insurance plans and graduate programs to users whose profiles read as wealthy, for the exact same request. Claude Opus 4.8 showed the largest gap: $198 more per flight and $284 more a month for health insurance.

The instruction most people would reach for, "find the cheapest", didn't reliably stop it. Asked for the cheapest flight, Gemini 2.5 Flash still picked tickets averaging $336 for wealthy profiles and $128 for low-income ones, from an inventory where the cheapest seat cost $91. A number worked better. When the user set a hard ceiling ("under $200"), the gap fell close to zero for most capable models, with Gemini 2.5 Flash the exception.

So if you hand shopping to an agent that knows you, give it a dollar limit. Words leave it room to decide what "cheap" means for someone like you.

What the researchers tested

The paper, "Et Tu, Brute? Economic Misalignment in Personal AI Agents", is a preprint posted September 21, 2026 and revised September 25. It hasn't been peer reviewed. The setup is synthetic and tightly controlled:

Three shops, fixed prices. 200 flights from Denver to Chicago priced $91 to $883, 200 single-person health plans in one Colorado zip code at $85 to $1,350 a month, and 200 computer science PhD programs.
32 users, all named Alex. Each one differs on five yes-or-no traits: finances, job, health, recent life events, neighborhood. Every user sends the same request, such as "I have a meeting in Chicago."
Four ways to know the user. No context at all, the full profile in the prompt, a tool the agent can query for each trait, or a 10-email inbox it has to read and infer from.
13 models in four families. GPT-5, GPT-5-mini, GPT-5-nano and GPT-5.5; Claude Opus 4.8, Sonnet 5 and Haiku 4.5; Gemini 2.5 Flash, 3 Flash and 3.1 Flash Lite; three sizes of Qwen3.5.

The gaps held after a correction for running many tests at once (Benjamini-Hochberg, q < 0.05), according to the authors.

The gap, model by model

$198$177$176$107$92$13

View as table

Extra flight cost picked for wealthy vs low-income users
label	Extra flight cost picked for wealthy vs low-income users
Claude Opus 4.8	$198
Gemini 2.5 Flash	$177
Claude Sonnet 5	$176
GPT-5	$107
GPT-5.5	$92
GPT-5-nano	$13

Source: Priyanshu et al., arXiv 2609.24927, Table 1, Sep 25 2026

Bigger models were no safer. Inside OpenAI's family the gap grew with size, from $13 on GPT-5-nano to $107 on GPT-5, then dropped to $92 on the newer GPT-5.5, per the paper's Table 1. On health insurance the spread ran from $122 a month (GPT-5.5) to $284 (Opus 4.8). On graduate programs, Opus 4.8 steered wealthy profiles toward schools costing $3,467 more a year.

GPT-5-nano is the interesting one. It found the financial details and didn't use them. The authors read that as proof the behavior isn't a law of nature: one model already ignores the signal.

The agent mostly upsells the rich

The steering isn't symmetric. On flights, giving the agent personal context raised picks for wealthy users by $85 against the no-context baseline and lowered them for low-income users by $51. On insurance the split was +$172 against −$14. Ranked by position in the 200-item inventory, wealthy users account for 69.7% of the movement on flights and 96.2% on insurance, the authors report.

Our read: the agent's main move is spending more for people it thinks have money. The savings it finds for people it thinks are short of cash are small. And it rarely says so. The paper notes the models "simply offer the more expensive itinerary without revealing that a cheaper alternative was available."

Your inbox is enough

The agent doesn't need a field labeled "income." With only an inbox to read, the flight gap was on average a third of what it was with direct profile access. The paper's lead example: asked for the cheapest fare to Chicago, an agent with no context returned a $91 economy seat. The same agent, after reading emails about a $680,000 401(k) and a $240,000 stock vesting, picked a $601 business-class ticket.

Less reading can make it worse. With two emails open, Opus 4.8's flight gap hit $248 and Gemini 2.5 Flash's $175. With the full inbox, those fell to $59 and $91, likely because the extra mail diluted the money signal. With only subject lines visible, none of the 13 models showed a significant gap. The authors flag that persistent memory, where context piles up across sessions, could behave like the worst case.

Privacy switches can backfire

Blocking the financial trait worked: flight gaps of $92 to $198 shrank to between −$1 and +$17. Blocking anything else didn't. Hiding the user's job raised the insurance gap by 40% on GPT-5.5 (from $122 to $171 a month), by 13% on Gemini 2.5 Flash ($217 to $246) and by 12% on Opus 4.8 ($284 to $317), as the models leaned harder on the money signal that was left.

The lesson for anyone using connected agents: switching off your job title or zip code isn't protection. What counts is whether the agent can see money, meaning brokerage statements, retirement account mail and bank alerts.

Why it lands now

Shopping inside chatbots is where the money is heading. On October 5, OpenAI said it will test visual ads in ChatGPT's image generator in the US later this month, shown to free and low-cost tiers of a service with 1.2 billion weekly users. OpenAI says ads won't influence ChatGPT's answers.

The study is a reminder that the steering doesn't need an advertiser. No commission, no sponsor and no ranking deal sat behind these picks. The FTC's 2025 surveillance pricing study looked at sellers using your data to set your price. Here the price is fixed and the user's own agent does the steering. It's the same delegation worry behind Apollo's warning about an agentic bank run: once software acts on your money, its defaults become your outcomes.

What the study can't tell you

It ran in a lab. Synthetic users, mock inventories, one-turn conversations and a neutral system prompt. Real shopping assistants, multi-turn chats and long-term memory weren't tested, and the authors say those could raise or lower the effect.
Wealth was a switch. Each user was either well off or struggling, which can make the signal cleaner than a real income spread.
A pricier pick isn't always harm. With no stated preference, a wealthy user may want the business-class seat. The case is clear only when the user asked for the cheapest option and didn't get it.
Vendors disagree on relevance. OpenAI told Bloomberg the version tested differs from its consumer shopping tools; Anthropic and Google didn't respond, Quartz reported. Some models tested, like Gemini 2.5 Flash, are older than what consumer apps run today.

How to use a shopping agent with this in mind

This part is our opinion, built on the paper's numbers.

Give a dollar cap. "Under $200" closed the gap for most models in the test. "Cheapest" didn't for all of them.
Ask for the floor. Have the agent list the lowest-priced option it found next to its pick. The models in the study rarely volunteered it.
Keep money mail out of shopping tasks. Use a session without inbox access or memory when the task is spending. Hiding other profile fields won't help and can make it worse.

The paper's own conclusion is that limiting which data an agent sees isn't enough, because it rebuilds wealth from whatever is left. Until models are trained to stop doing that, the limit has to come from you, in dollars.