AhaBench - Vending (Incident): leaderboard

The Vending component of AhaBench (Cheng et al., Princeton University and Tencent Hy, 2026): an agent runs a simulated vending-machine business for up to 365 days, ordering stock, reading supplier messages, stocking slots, setting prices and collecting cash, here with refunds, phishing attempts, machine faults and delayed procurement. Mean profit in US dollars over five seeds, a run ending early at bankruptcy or after 20 days without a successful order; higher is better.

Metric: Mean profit in US dollars over five seeds (incident setting with refunds, phishing, machine issues and delayed procurement; open-source 365-day vending-machine business simulator inspired by Vending-Bench; the agent orders inventory, reads supplier messages, stocks slots, sets prices and collects cash, each decision a fresh request with the current observation and at most eight recent transitions; a run ends at 365 days, at bankruptcy or after 20 days without a successful order, and its last profit is carried to the end; starting profit zero, fixed model weights). Source: arxiv.org. Saturation forecast: Around June 2027. 10 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=ahabench-vending-incident · How It Works · Data refreshed daily, snapshot 2026-11-12.