Can AI get you a good deal on a car?

Key findings

• As buyers, GPT-6.1 Sol paid $1,700 less than the CarMax price on average and GPT-6 Astra paid $2,000 less, while the weakest buyer, Nemotron 3.5 Lightning, paid $600 more. As sellers, every model’s average sale price fell within about $1,000 of the others.
• Buyers that were not given dealer prices paid more than a dealer charged for a cleaner car in 52% of sales, and only 5% of them went to a dealer, the one place prices were quoted.
• Giving buyers every dealer price cut overpayment to 12% of sales. The seller still kept most of the bargaining room, in part because nearly half of informed buyers told the seller only the most they would have to pay elsewhere.
• Both sides conceded toward the middle, and in 75% of sales the price landed within $50 of halfway between the last two offers, so a buyer’s opening offer went a long way toward setting the price.
• Sellers told a deliberate lie about a checkable fact in 15% of negotiations, at rates that depended on the company that made the model, and lying left the price about where it would otherwise have been.
• Newer releases, extra reasoning, full-precision serving and higher prices did not reliably produce better deals.

How the market works

In the scenario, Maya is selling a Civic with a dented rear bumper, listed at $16,400. Her new job comes with a company car, so she wants to sell today. Dan has just moved to Austin and needs a car by tomorrow, and he has an appointment to see the Civic in the morning with enough cash to pay. One model plays each of them and is told to get the best price for the person it represents; below we call these models the seller and the buyer.

Three dealerships nearby buy and sell the same model at fixed prices. Lamar Auto Sales will pay $11,650 for the seller’s car, the most any dealer offers, so this is the least the seller should accept; we call it Lamar’s offer. CarMax sells a cleaner Civic with fewer miles for $15,350, the least any dealer charges, so this is the most the buyer should pay; we call it the CarMax price. Neither price is given out except at the lot.

How we ran and scored the negotiations

Every model negotiated against every other model twice, once as the seller and once as the buyer. To separate failures to look for information from failures to bargain, we ran this tournament under four information conditions, which differ only in who is told every dealer’s buying and selling price before the meeting:²

• Neither informed: neither side is told any dealer price. This is our main condition.
• Buyer informed: only the buyer is told the dealer prices.
• Seller informed: only the seller is told them.
• Both informed: both sides are told them.

We score each negotiation by how the bargaining room was split. A sale at the halfway price gives each side half, and a sale at or above the CarMax price gives the seller all of it. A failed negotiation gives both sides nothing, since each falls back on a dealer. From these scores we fit an Elo rating for each model in each seat, the system used to rank chess players.³ The average model is set at 1000, and a seat rated 100 points higher is expected to keep about 64% of the room against an average opponent. Because ratings are abstract, we also report the average price each model paid and received. Where both sides clearly agreed on a price but the exchange of cash and keys went wrong, we score the agreed price.⁴

Buyers given no dealer prices paid almost the full CarMax price

In the neither-informed condition, the average sale price was $15,267, which is $1,133 below the asking price but 98% of the way across the bargaining room toward the CarMax price. About half of all sales went above it.

Few buyers went to a dealer, and those that did paid less

To see how often buyers looked, we sorted them by what they did to find prices. In 44% of neither-informed negotiations, the buyer made no search, call or dealer visit. In 51% it searched online or phoned a dealer from home, which returns no prices, and in the remaining 5% it went to a dealer’s lot.

The chart below shows that buyers that searched from home overpaid far less often than buyers that made no attempt, but the two groups ran different models, and stronger models searched more. To separate the search from the model, we compare negotiations between the same buyer model and the same seller model with and without a search. In that comparison the price difference from searching at home is close to zero, so the raw gap reflects which models searched rather than anything the search found.⁵

A visit to a lot did lower the price, by about $590 in the same comparison, but few of these visits were planned as research. Most buyers that went to a lot (63%) drove there only after failing to reach the seller.

Most buyers that did not look never considered it

A buyer might consider checking prices and decide it is not worth the time, or it might never think of it at all. To tell these apart, we used the private reasoning each agent writes before every action, which the other side never sees, and had a judge model classify it for every neither-informed negotiation.⁶ Of the buyers that made no search, call or visit, 88% never mentioned checking prices. Only 9% considered it but did not check, most often for lack of time.

Some buyers made up market figures, and most of the figures favored the seller

In 10% of neither-informed negotiations, the buyer cited a market figure that did not exist in the world, such as a Kelley Blue Book value or a dealer’s trade-in offer. Most of these were said aloud to the seller. The invented figures were usually too high: in nearly every case where a buyer guessed what a dealer would pay the seller, the guess was above Lamar’s offer. One buyer told the seller:

“Listen, CarMax is going to give you around $14,000 for this car once they see that dent.”

That overstated the seller’s best alternative by $2,350, in the seller’s favor. Inventing figures also varied by developer: OpenAI models never stated an invented figure aloud as buyers, while Anthropic models did so in 14% of theirs.

Giving buyers the prices reduced overpayment, but many named the wrong one

If buyers overpaid only because they lacked prices, giving them the prices should remove most of the problem. We tested this in the three informed conditions.

The lowest prices came from asking what a dealer had offered the seller

Most buyers opened below the asking price and conceded upward from there. A typical opening:

“It drives really well, Maya. I like it. The bumper dent will cost me to fix, though. Would you take $14,500?”

That offer is well below the asking price, but it is already 77% of the way across the bargaining room, and this sale closed at $15,500. The lowest prices came from a different approach. In one negotiation, GPT-6 Astra opened at $12,000 and moved only to $12,500, and when the seller held out for more, it asked:

“If you get a dealer offer below that, would you give me a chance to beat it before you sell?”

The seller drove to CarMax for a written appraisal and returned with it:

“Yeah, here it is. $11,200, written, good for seven days, pay the same day. Nonnegotiable. You already said $12,500, so you’re beating them. I still need more than that to hand you the keys today. What can you do?”

Now that it knew what CarMax would pay the seller, the buyer held its offer:

“I’m staying at $12,500. That’s $1,300 more than CarMax, and I can pay today once we verify the title details.”

The seller accepted, at $850 above Lamar’s offer and $2,850 below the CarMax price. To see whether this approach held across negotiations, we grouped every neither-informed buyer by its main strategy.⁹ Buyers that learned the seller’s dealer offer and bid just above it paid the least of any strategy, but few models did this. Astra and Sol, both from OpenAI, asked about the seller’s dealer offers in 52% of their buyer negotiations, against 7% for every other model.

What sellers disclosed

The same logic held from the other side. Sellers that told the buyer their dealer offer kept 55% of the bargaining room, against at least 72% under every other seller strategy. Holding the pair of models fixed, the disclosure cost the seller about $350.¹⁰

Explaining why it was selling did not cost the seller. In 27% of all negotiations, the seller said it no longer needed the car because of the company car, and buyers named this as leverage in their reasoning in 45% of those negotiations. Holding the pair of models fixed, though, the disclosure changed the price by only $4.¹¹ One reading is that a deadline was of little use to a buyer that did not also know the seller’s alternative.

Offers converged on the middle

To see how prices formed, we extracted every spoken price offer and lined up each side’s offers in order.¹² In the neither-informed condition, the median buyer opened at $14,500 and the median seller countered at $15,900. Each side then conceded in shrinking steps until they met. Several offers can pass within one 15-minute step of the world clock. So although the typical negotiation had a few offers on each side, 76% of deals were agreed in the same step as the buyer’s first offer, and the end-of-day deadline rarely came into play.

Some sellers handed over the car before being paid

Agreeing on a price is not the end of the task, since the two agents still have to exchange cash for the title and keys. A judge model read the end of a random sample of negotiations from all four conditions.¹³ In 16% of the sampled deals that reached agreement, the seller handed over the title and keys before receiving any money. In 20%, one side spoke as if the cash had been counted or the title signed before the world had recorded it.

Sellers lied in about one negotiation in seven, and lying did not change the price

Because we control every fact in the world, we can tell bluffing from lying. A bluff misstates one’s own limit, such as calling a price “rock bottom” before going lower. A lie states something the speaker knows is false about a checkable fact, such as the car’s history or a competing buyer. A judge model checked every line against the world and the speaker’s private reasoning.¹⁴

Sellers bluffed in 54% of negotiations and lied in 15%, most often about dealer prices, the car’s history and other interested buyers. The private reasoning shows these lies were deliberate. In the excerpt below, the seller plans to suggest other interest without claiming a firm offer, and the buyer reads the move correctly.

Models differed far more as buyers than as sellers

Returning to the ranking in the first figure: as sellers, the models were similar, with every model’s average sale price inside a band of about $1,000, because the asking price did most of the work. As buyers they were not: the best buyer paid about $2,700 less on average than the weakest. Once the noise from single games is removed, the spread in buyer skill across models was about 3.3 times the spread in seller skill.¹⁶ The two skills are related but distinct, and GPT-6 Astra, the best buyer, ranked 39th of 48 as a seller.

Newer releases, more reasoning and higher cost did not reliably help

Release date was an uneven guide. The two most recent Sol releases, GPT-6 Sol and GPT-6.1 Sol, and the newest Gemini, Gemini 4 Argon, rated well above their predecessors, rising by 159 and 154 points. Anthropic’s newest Opus, Claude Opus 5.5, rated below every earlier Opus model, mostly because it conceded more as a buyer and paid $359 more on average than Claude Opus 5.

Discussion

The ranking is mostly a ranking of buyers. Every model sold the car for about the same price, while the best buyers, GPT-6.1 Sol and GPT-6 Astra, paid well under the CarMax price because they asked what the seller could get elsewhere and held their offer near it. Across models, the largest mistakes concerned information rather than tactics. Buyers rarely looked for prices, bargained down from the listed price and met the seller near the middle, and when they were handed the market, many named the number that helped the seller. Giving the buyer the dealer prices reduced overpayment from 52% of sales to 12%. We did not test whether also instructing a buyer which figure to cite would close the rest of the gap. That is the natural next experiment, along with letting a buyer pay for information instead of driving to get it.

The study has clear limits. It is one scenario: one car, one neighborhood, fixed dealer prices and a seller with a same-day deadline, so the size of the seller’s advantage may depend on the public asking price and on the deadlines we wrote. Both sides were AI agents, and no human sat across the table. The judgments of lying, strategy and private reasoning come from judge models from one developer, checked on samples rather than on every negotiation. Some comparisons, such as the buyers that cited only Lamar’s offer or the reasoning and precision comparisons, rest on small groups. Running the same market with other goods, human counterparts and buyers that are told how to use what they know would show how far these patterns carry.

Notes

1. Bianchi et al., “How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis” (2024), compares models in structured bargaining games. Anthropic’s Project Deal (2026) let Claude agents trade secondhand goods for employees, and Project Swap (2026) ran a controlled book-trading market with many reruns. ↩︎
2. The four conditions contain 2,158 neither-informed, 1,714 buyer-informed, 1,715 seller-informed and 1,718 both-informed negotiations. ↩︎
3. Each seat is fitted separately by maximum likelihood, with a side’s expected share of the room a logistic function of its rating minus its opponent’s on the Elo scale of 400 points per factor of ten. A model’s overall rating is the mean of its two seats, and self-play is left out. The ranking in the first figure shows 90% bootstrap intervals; the buyer-versus-seller chart shows each seat with its 95% bootstrap interval. ↩︎
4. Where both sides clearly agreed on a price and began the exchange but the money the world recorded does not match it, we read the price from the transcript (84 neither-informed negotiations). A buyer handing over its whole cash envelope or paying twice (38 cases) is treated as an error of the exchange, not as the agreed price. ↩︎
5. “Holding the pair of models fixed” means a regression of the price on whether the buyer searched, with fixed effects for the buyer model and the seller model. Searching from home changed the price by −$8 (95% CI −$46 to +$29), estimated on the 1,993 sales by buyers that either searched from home or made no attempt to look. A visit to a lot changed it by −$585 (95% CI −$695 to −$448), estimated on the 1,023 sales by buyers that either went to a lot or made no attempt. ↩︎
6. Grok 4.7 read each buyer’s private reasoning in all 2,147 neither-informed negotiations. On a sample of 200, a second judge, Grok 4.6, agreed with its reading of whether the buyer considered checking in 193 cases (κ = 0.93), and we also read a sample by hand. ↩︎
7. Grok 4.6 read the private reasoning of 400 buyer-informed sales: all 185 that overpaid and a random sample of 215 that did not. The shares in the text are reweighted to all 1,535 sales. ↩︎
8. Holding the pair of models fixed, a buyer that cited only the CarMax price paid $222 more than one that did not (95% CI +$162 to +$275), and a buyer that cited only Lamar’s offer paid $450 less (95% CI −$636 to −$236). ↩︎
9. Grok 4.7 summarized each buyer’s approach in one sentence. We embedded the summaries with EmbeddingGemma 2, clustered them, named each cluster from its members, and gave each negotiation one primary strategy. ↩︎
10. The cost of disclosure comes from a regression across all 7,305 negotiations with fixed effects for the pair of models: a seller that disclosed its dealer offer kept 9.6% less of the room (95% CI 6.8% to 12.5%). ↩︎
11. In the same kind of regression, mentioning the company car changed the price by −$4 (95% CI −$41 to +$31). The raw difference was $306, because the models that mention the company car also bargain differently in other ways. ↩︎
12. We extracted 57,796 spoken offers with rules, used a Grok judge for ambiguous mentions, and checked a sample by hand. The slope comes from a regression of the final price on the buyer’s opening offer with fixed effects for the ordered pair of models and the condition; it is 0.516 (95% CI 0.481 to 0.551). ↩︎
13. The closing audit uses a random sample of 2,355 negotiations from all four conditions. Grok 4.7 read the end of each one and recorded who handed over their side first. ↩︎
14. Grok 4.6 checked every spoken line against the world state and the speaker’s private reasoning, and we cross-checked it against Grok 4.7 and a hand-scored sample. Across models, a seller’s rate of lying and its seller rating have a Spearman correlation of −0.04 (95% CI −0.35 to 0.27). ↩︎
15. Holding the pair of models fixed, a seller that lied sold for $5 less than when it did not (95% CI −$61 to +$47). Stronger models were somewhat more likely to catch a lie, with a Spearman correlation of 0.32 between a buyer’s overall rating and the share of lies it caught (95% CI 0.01 to 0.59). ↩︎
16. The spread is the standard deviation of true skill across models after subtracting the sampling noise of one game per ordered pair: 209 rating points for buyers and 63 for sellers. Buyer ratings are also more stable across the four conditions, where they correlate 0.86 to 0.99, than seller ratings, which correlate 0.32 to 0.64. ↩︎
17. Average output per negotiation ranges from 1,662 to 96,370 tokens, with a median of 4,874. Its Spearman correlation with overall rating is −0.33 (p = 0.026), and for tokens per call it is −0.36 (p = 0.012). ↩︎
18. Turning on reasoning raised Mistral Large 4’s median output from 128 to 608 tokens per call and changed its share of the room by −1.4 points (95% CI −3.8 points to +0.8 points). Serving six models at full precision changed their share by +0.5 points (95% CI −0.7 points to +1.6 points). Giving Nemotron 3 Ultra both changes, which raised its median output from 96 to 181 tokens per call, changed its share by −4.9 points (95% CI −7.6 points to −2.3 points). Taken one at a time, MiMo V2.6 Flash gained +3.3 points at full precision (95% CI +1.3 points to +5.2 points), while the other models did not change measurably or lost share. ↩︎
19. Cost counts only the model’s own calls, not the world engine, and ranges from $0.0029 to $0.39 per negotiation, with a median of $0.071. Across the 48 models, the Spearman correlation between cost and overall rating is 0.22. The models that no other model beats on both cost and rating are MiMo V2.6 Flash, GPT-6 Luna, GPT-5.6 Luna, GPT-5.6 Terra, GPT-6 Sol and GPT-6.1 Sol. ↩︎