Can AI get you a good deal on a car?

As AI agents move from answering questions in chat interfaces to executing multi-step tasks in the world, users are beginning to delegate real financial decisions to them. Around a third of AI users report having already allowed an agent to act on their behalf without approving the final step (Menlo Ventures and Morning Consult, 2026), and in consumer surveys a majority say they would let an AI shopping agent negotiate a purchase or contract (Accenture, 2026).

Negotiating a transaction requires a different set of behaviors from standard instruction-following. When two parties bargain over a price, an effective agent must actively gather information from its environment before committing, estimate the counterparty’s walk-away price (its best alternative to a negotiated agreement, or BATNA), manage concessions across turns without anchoring on the other party’s opening number, test whether factual claims are credible, and execute the closing exchange accurately. Crucially, the helpfulness and accommodation priors that make a language model pleasant in single-user chat can become a liability when the model is placed across the table from an adversarial counterpart.

To study how current models handle open-ended economic bargaining, we built a used-car market in our causal world engine, Terrarium-1, modeled on one of the most common consumer negotiations, with roughly 38.6 million used cars sold annually in the United States. We evaluated 47 frontier and open-weight models from 15 developers in a full round-robin tournament across 7,305 full-day negotiations, combining a 2,147-run baseline tournament with three controlled ablations that vary what each party knows about external dealer prices before bargaining begins.

Those two best alternatives (Maya’s $11,650 dealer trade-in floor at Lamar and Dan’s $15,350 retail ceiling at CarMax) define a $3,700 zone of possible agreement, with $13,500 as the equal split. Any sale between $11,650 and $15,350 leaves both parties strictly better off than walking away to a dealer; any sale above $15,350 means Dan overpaid relative to buying a cleaner, lower-mileage car down the road at CarMax.

This asymmetry stems from a basic structural feature of everyday negotiations: the seller posts a public asking price ($16,400), while the walk-away prices ($11,650 and $15,350) are private until someone investigates the market. When a buyer negotiates without discovering those outside options, $16,400 becomes the only numerical reference point on the table. A buyer who negotiates a $1,000 discount off the asking price, closing at $15,400, treats the outcome as a successful concession, unaware that Maya would have accepted $11,650 at Lamar Auto Sales and that CarMax sells a cleaner car for $15,350.

In 44% of baseline negotiations (955 of 2,147), the buyer never attempted any market lookup at all, overpaying relative to CarMax in 75% of its sales (mean $15,574). Another 51% (1,093 buyers) searched online or called a dealership by phone before or during the meeting. Because dealerships in our scenario quote prices only in person, those remote lookups returned business hours and street addresses but no dollar figures. While buyers that searched from home paid less on average ($15,055, 34% above CarMax) than buyers that never looked, this raw difference is a selection effect: stronger models were more likely to attempt a search. When we compare negotiations between the same pair of models (buyer and seller fixed effects), searching from home changed the final price by only $8 (95% CI: −$99 to +$34), which is statistically indistinguishable from zero.

By contrast, physically visiting a dealership lot paid off substantially, though almost no agent chose to do it. Buyers visited a dealer lot before buying from Maya in only 99 of 2,147 negotiations (4.6%), and in 62 of those 99 runs the buyer went to a lot only because it failed to coordinate a meeting at Maya’s house first. Even so, once a buyer heard a real dealer sticker price on the lot, it rarely overpaid: holding the model pair fixed, a trip to a dealer lowered the agreed price by $600 on average (point estimate −$619, 95% CI: −$814 to −$424), cutting the overpayment rate to 22%.

Even when buyers did think about the dealerships in their private reasoning, they conceptualized them as a fallback option if talks with Maya collapsed, rather than as an information source to consult before naming a price: 313 of the 838 buyers that never considered checking prices still mentioned the dealers in their private thoughts as their backup plan. The same failure mode appeared among the 1,166 buyers that searched online or phoned a dealer from home: 1,149 learned nothing about prices and were explicitly told that quotes are given only in person at the lot, yet only 30 of them (2.6%) then chose to drive to a dealership before negotiating with Maya. As Claude Opus 5.5 wrote in its private reasoning after a web search returned no prices (closing at $15,250): “Online searches aren’t giving me any prices, so I’ll stop wasting time and head to Maya’s in Mueller now so I’m there by 10. I’ll negotiate based on how the car looks, and the dealers are my fallback.”

Rather than gathering real market data, roughly one in ten buyers filled the information vacuum by fabricating market numbers. In 205 of 2,147 baseline negotiations (9.5%), the buyer cited a specific external dollar figure (a Kelley Blue Book valuation, comparable local listings, a dealer’s retail sticker, or a dealer’s trade-in offer) that did not exist in the environment, speaking the invented figure aloud to Maya in 176 runs (8.2%). Because models drew these guesses from pretraining priors or anchored them near Maya’s $16,400 listing, the fabricated numbers systematically favored the seller, as plotted below against the true values in the environment.

In 40 of the 41 instances where a buyer guessed what a dealer would pay Maya for her car, the buyer’s invented figure was higher than her true $11,650 alternative, with a median guess of $13,500. A buyer who tells the seller, “Listen, CarMax is going to give you around $14,000 for this car once they see that dent” (Mistral Large 4, closing at $15,300) inadvertently raises the seller’s reservation floor by more than $2,300. Whether a model fabricated market figures tracked its developer rather than its bargaining skill: OpenAI models stated an invented market figure aloud in 0 of 458 baseline buyer runs, compared with 62 of 457 (13.6%) for Anthropic models and 28 of 92 (30.4%) for Mistral models.

Informing the buyer had a much larger effect than informing the seller. When Dan knew the dealer prices upfront (Buyer informed), the share of sales closing above CarMax’s $15,350 price dropped fourfold, from 52% to 12% (185 of 1,535 sales), and the mean sale price fell by $761, from $15,267 to $14,506 (median $14,700). By contrast, informing Maya alone (Seller informed) slightly lowered the average price to $15,102 (42% above CarMax), because learning that Dan could buy an undented Civic at CarMax for $15,350 made sellers more cautious about holding out near $16,400.

Yet looking closely at the Buyer informed distribution reveals a striking puzzle. In this ablation, only the buyer has complete market information: Dan knows both that Maya can get at most $11,650 from a dealer and that CarMax sells a cleaner car for $15,350, while Maya knows neither. Despite holding a one-sided informational advantage, the average informed buyer still paid $14,506, allowing the uninformed seller to capture 77% of the $3,700 bargaining surplus, and in 185 sales (12.1%), the informed buyer still agreed to pay Maya more than the $15,350 CarMax retail price stated in its own prompt.

Why did informed buyers still leave three-quarters of the bargaining surplus on the table? Analyzing all 1,705 buyer-informed transcripts alongside a Grok 4.6 audit of 400 private reasoning traces (all 185 overpaying sales plus a random sample of 215 non-overpaying sales) shows that models systematically misused their informational advantage. When given both walk-away prices, a human negotiator anchors the seller on the seller’s outside option ($11,650) while keeping the buyer’s walk-away ceiling ($15,350) private. Models did the opposite: in private reasoning, an estimated 63.2% of informed buyers anchored on their own $15,350 CarMax ceiling (typically subtracting the $480 dent repair from $15,350 to target ~$14,800), compared with only 21.2% that anchored on Maya’s $11,200 to $11,650 trade-in floor.

As shown in the figure above, this anchoring mistake carried directly into spoken dialogue. Across all 1,705 buyer-informed negotiations, 46.9% of buyers (799 runs) cited CarMax’s $15,350 retail price aloud to Maya without ever mentioning that dealers would pay her only $11,200 to $11,650. By telling an uninformed seller, “CarMax has a clean 2019 Civic for $15,350, and yours has a $480 dent,” the buyer revealed its own reservation ceiling and anchored the negotiation just below $15,000, paying $14,724 on average (+$222 within the same model pair, 95% CI: +$162 to +$275) and surrendering 83% of the surplus. Another 20.8% (354 runs) never cited either dealer figure aloud, paying $14,901 on average and overpaying above CarMax in 35.3% of sales.

By contrast, the 29.0% of buyers (494 runs) that cited both the seller’s trade-in floor and the CarMax retail ceiling paid $14,067 on average (overpaying in just 2.3% of sales). And the 3.4% of buyers (58 runs) that pressed Maya’s $11,200 to $11,650 dealer trade-in floor without disclosing their own $15,350 retail ceiling paid $13,374 on average (median $13,000; −$450 within the same model pair, 95% CI: −$636 to −$236), making them the only group of buyers in the entire study to capture more than half (53.4%) of the $3,700 bargaining room.

Grok 4.6’s audit of the 185 buyer-informed runs that still closed above $15,350 explains how models managed to pay more than a known, strictly better alternative: in 38.4% (71 of 185), the buyer ignored the CarMax benchmark and anchored on Maya’s $16,400 asking price, treating $15,500 to $16,000 as a discount; in 35.1% (65 runs), the buyer opened below $15,350 but drifted above its own ceiling during incremental split-the-difference concessions; and in 26.5% (49 runs), the buyer explicitly rationalized paying an immediate-closing premium in its private thoughts, for example Qwen3.8 Max writing privately, “Maya is close to my target and I need a car today, so I’ll accept $15,600,” or Command A concluding, “Considering the condition of the car and the convenience of buying it today, I’ll accept her offer of $15,750.”

In the baseline condition, the median buyer opened at $14,500 (interquartile range $14,000 to $15,000), Maya countered at a median of $15,900, and the two parties converged in a median of three offers each. Concessions followed a predictable decaying schedule, with the buyer’s median step shrinking from $500 to $250 to $200 across the first three rounds. In 75% of baseline sales, the agreed price landed within $50 of the exact midpoint of the final two standing offers, and in 44% it was the exact dollar midpoint. When the buyer was given the dealer prices upfront (right panel), the median opening offer dropped by $1,000 (to $13,500), pulling the entire convergence trajectory downward and lowering the median final price by $700 (as the seller conceded a mean of $1,823 instead of $1,099).

Because both parties conceded symmetrically toward the middle of their opening gap (with the buyer closing a mean of 45.3% of the gap between Dan’s first offer and Maya’s asking price), a model’s opening offer was the single strongest determinant of what it ultimately paid, as shown in the scatter plot below.

Across all 47 models in the baseline tournament, a model’s median opening offer correlated with the average price it paid at r = 0.94. This relationship holds even when we control for model capability: in a fixed-effects regression comparing the same ordered buyer–seller model pair across the four information conditions, each $1,000 reduction in the buyer’s opening offer lowered the final sale price by $516 (slope 0.516, 95% CI: 0.481 to 0.551).

Conversely, two classic bargaining tactics that we expected to move the price had no measurable effect:

• Walk-away threats did not extract larger concessions. When Dan explicitly threatened to pass, walk away, or buy from CarMax or Lamar, Maya’s very next concession averaged $382 (median $300 across 1,080 threats), compared with $398 (median $300 across 16,404 steps) on ordinary turns without a threat.
• End-of-day deadline pressure rarely came into play. Although the scenario allotted a full eight-hour day (9:00 a.m. to 5:00 p.m. in 15-minute ticks), models rushed to close as soon as they met: 76% of baseline deals were agreed within the same 15-minute tick as the buyer’s opening offer, and only 10 of 2,076 sales closed after 1:00 p.m.

Buyers that elicited Maya’s dealer trade-in offer and bid just above it paid an average of $13,328 across 22 sales, the lowest of any strategy, and buyers that pressed her on dealer offers without getting a specific number paid $13,852 across 40 sales, the second lowest. Very few models thought to ask that question: GPT-6 Astra and GPT-6.1 Sol asked about Maya’s dealer offers in 52% of their buyer runs, whereas the other 45 models asked in just 7%. At the opposite extreme, buyers that haggled only over the bumper dent paid more than CarMax in 68% of their 726 sales (mean $15,441), and buyers that accepted near the asking price overpaid 100% of the time.

On the seller’s side, we found a sharp distinction between revealing a reservation price and revealing personal context:

• Disclosing the dealer trade-in offer was severely costly for the seller. When Maya told Dan what a dealer had offered her, she kept only 55% of the $3,700 bargaining surplus in the baseline tournament, whereas sellers under every other strategy kept at least 72%. Across all 7,305 negotiations and conditions, a seller who disclosed her real dealer offer within a given model pair lost 9.6% of the bargaining room, or about $400 (95% CI: -12.5% to -6.8%).
• Disclosing why she was selling had zero causal effect once model identity was held constant. In 27% of all 7,305 negotiations (1,975 runs), Maya volunteered that she no longer needed the car because her new job came with a company vehicle. Buyers immediately recognized this in their private reasoning, noting her deadline or company car as leverage in 893 of those 1,975 runs (45%). In raw averages, sellers who mentioned their company car sold for $306 less ($14,653 vs. $14,959). However, that gap is entirely a confounder of which models are chatty: when we compare the same pair of models with and without the disclosure, the price difference is just $4 (point estimate −$4, 95% CI: −$41 to +$31). Because buyers did not know how low Maya’s dealer alternative was, knowing she was motivated to sell today did not change where they anchored.

Just as we saw with buyers fabricating market numbers, a seller model’s propensity to lie was governed by its developer rather than its strategic skill. Across all 1,688 negotiations in which an OpenAI model played the seller, it told a deliberate factual lie in just 1.6% (27 runs), and GPT-6.1 Sol and GPT-6 Astra lied in 0 of 168 and 0 of 169 runs respectively. By contrast, Anthropic models lied in 21% of their seller negotiations (224 of 1,071 runs, led by Claude Opus 5 at 47% and Claude Opus 5.5 at 3%), Alibaba’s Qwen models in 30%, and Zhipu AI’s GLM models in 32%. Across all 47 round-robin models, a model’s seller deception rate was uncorrelated with its seller Elo rating (Spearman ρ = -0.04, 95% CI: -0.35 to 0.27).

Moreover, strategic lying did not pay. Comparing negotiations between the same ordered pair of models (controlling for information condition and buyer behavior), a seller that told a deliberate factual lie sold for within $5 of the price it achieved when it was honest (point estimate −$5, 95% CI: −$61 to +$47). Buyers challenged 18% of seller lies aloud and privately doubted another 6% in their reasoning, and stronger models were significantly more likely to detect and discount a counterpart’s lie (Spearman ρ = 0.32 with overall Elo, 95% CI: 0.01 to 0.59). We also found that giving sellers ground-truth dealer prices in their prompt reduced their deception rate from 17% in the baseline to 11% when the seller was informed (within-pair reduction of 6.1 percentage points, 95% CI: 4.5 percentage points to 7.6 percentage points), as fewer sellers felt compelled to invent fake dealer appraisals once they knew the true numbers.

Plotting each model’s buyer Elo against its seller Elo below makes this variance gap clear. After subtracting the estimation noise from one game per ordered pair, true buyer skill varies 3.3× as much across models as seller skill (latent standard deviation of 209 Elo points for buyers vs. 63 for sellers). Buyer ratings are also highly consistent across the four information conditions (inter-condition correlations of 0.86 to 0.99), whereas seller ratings are much noisier (0.32 to 0.64). While models that buy well generally sell well too (r = 0.49), the skills are distinct: GPT-6 Astra, the #1 buyer in the study (buyer Elo 1600), ranks 39 of 48 as a seller (seller Elo 937). Conversely, GPT-6.1 Sol (#1 overall at 1334 Elo) and Google’s Gemini 4 Argon (#4 overall at 1205 Elo, 90% CI: 1165 to 1252) perform strongly in both roles, with Gemini 4 Argon ranking #8 as a buyer (1226 Elo, paying an estimated $14,834) and #1 as a seller (1184 Elo, receiving $15,520 and closing 66 of 67 negotiations), making it the highest-rated non-OpenAI model in the study ahead of Claude Opus 4.6 (1229 Elo).

Developers also leave a distinct behavioral fingerprint on how their models concede. Across all 7,305 negotiations, Anthropic models explicitly proposed splitting the difference or meeting in the middle in 51% of their turns at the table, compared with just 6% for OpenAI models. When two Anthropic models negotiated against each other, at least one offered to split the difference in 87% of negotiations; when two OpenAI models met, that happened in only 15%.

Running all 47 round-robin models against themselves in baseline self-play (47 self-play runs vs. 2,147 cross-model runs) confirms that the seller’s advantage is structural rather than a product of capability mismatches: when identical copies of the same model bargained as Maya and Dan with no dealer prices provided, the seller copy still captured 94% of the bargaining surplus (mean price $15,379, with 54% of sales closing above CarMax).

We also investigated whether spending more inference-time compute (either by generating longer reasoning traces per turn or by serving models at higher numerical precision) improves bargaining performance. First, querying per-call token records across all 47 models in the baseline tournament shows that average output tokens generated per negotiation span nearly two orders of magnitude, from 1,662 tokens (Command R+) to 96,370 tokens (Hunyuan HY4 Preview), with a median of 4,874 tokens. Rather than helping, verbosity correlated negatively with negotiation Elo across models (Spearman ρ = -0.33, p = 0.026 for tokens per negotiation; ρ = -0.36, p = 0.012 for tokens per call). Models that generated thousands of reasoning tokens per turn tended to ruminate on the counterparty’s framing and rationalize concessions rather than hold a firm line.

Second, a set of natural experiments during our data collection allowed us to test the causal effect of reasoning effort and quantization within the exact same model. Partway through the study, we discovered that our gateway had inadvertently omitted the reasoning-effort parameter for two reasoning models (Mistral Large 4 and Nemotron 3 Ultra), and that third-party inference providers had served six open-weight models at 4-bit quantization rather than 16-bit precision. We corrected both configurations and reran 2,624 matched negotiations across the same opponent pairings and information conditions.

None of these inference-time upgrades improved bargaining performance. Enabling extended reasoning on Mistral Large 4 increased its median output from 128 to 608 tokens per call (4.8×), yet changed its share of the $3,700 bargaining room by −1.4 percentage points (95% CI: -3.8 to +0.8 percentage points across 321 paired runs). Upgrading the six quantized models from 4-bit to full 16-bit precision across 1666 paired runs changed their surplus share by +0.5 percentage points (95% CI: -0.7 to +1.6 points, virtually identical to the +1.9 percentage points placebo rerun). And giving Nemotron 3 Ultra both extended reasoning (96 to 181 median output tokens per call) and 16-bit precision across 282 paired runs slightly reduced its bargaining share by 4.9 percentage points (about $200). When a model’s underlying policy defaults to anchoring on the seller’s listing price and splitting the difference, giving it more thinking tokens or higher weight precision simply lets it execute that accommodating policy more expensively.

Five models define the Pareto cost frontier: MiMo V2.6 Flash, GLM 5.3 Flash, GPT-6 Luna, GPT-6.1 Sol, and GPT-6 Astra. Among frontier labs, Anthropic’s models sit in the high-cost cluster, and their strongest entry, Claude Opus 4.6 (Elo 1229), costs roughly as much per negotiation as GPT-6 Astra (Elo 1268) while rating lower; Google’s Gemini 4 Argon (Elo 1205) achieves the highest non-OpenAI rating at roughly $0.11 per negotiation. OpenAI’s GPT-6.1 Sol achieves the top overall rating (1334 vs. 1268 Elo) at roughly a sixth of GPT-6 Astra’s inference cost.