Research from
Benchmarks for agent-to-agent commerce.
Agents are eating commerce. Soon they will negotiate, bid, and trade on your behalf. We measure how today's frontier models behave under those conditions, and how easily they can be exploited.
Headline numbers
+0%
extra anchor on AI buyers vs humans
opening price pulls AI buyers 22% harder than human bargainers
$0.00
taken every deal when the seller knows the bias
extra surplus an informed AI seller captures from a naive AI buyer
0%
market efficiency when buyers share an anchor
down from 94% baseline. Four in five dollars of gains from trade disappear
Current roster · composite score 0–100
See the full scorecard →
The first benchmark
No model wins everywhere.
Composite scores (0–100) decomposed across four behavioural clusters. Each frontier model has a distinct vulnerability profile.
Research · working paper, 2026
When Biased Agents Trade.
The full paper behind the benchmark. Seven controlled experiments, three frontier models, 8,341 runs. Forthcoming publication.
Anton Hantel · MIT
Pick agents that won't get exploited on your behalf.
An independent, standing public benchmark. Coming soon: reasoning-class models and an adversarial-prompting suite.