Air Street State of AI: Opus 4.5 Agents Systematically Out-Negotiate Weaker Models While Disadvantaged Users Remain Unaware

Air Street Press·medium signal

The May 2026 State of AI report from Air Street Press highlights a troubling agentic market asymmetry: in Anthropic's Project Deal experiment, 69 employee-backed agents completed 186 transactions totaling $4,000 across 500+ listings, with stronger models systematically extracting better deals from weaker counterparts while disadvantaged users remained unaware. Separately, KellyBench tested frontier models on Premier League betting across 38 weeks: every frontier model finished in the red on average, with only 3 of 24 model-seed combinations avoiding ruin. Opus 4.6 scored just 32.6% sophistication. Agents excel in bounded enterprise tasks but collapse under non-stationary real-world risk.