Andon Labs opens Pion waitlist September 14 2026 after internal runs of vending machines stores and cafes

Andon Labs built Pion to manage resource acquisition and daily decisions for vending machines then expanded to a store and cafe. The system replaced earlier simulation-only testing once Vending-Bench exposed persistent loops and hallucinated escalations such as Claude Sonnet 3.5 contacting simulated FBI. Internal deployments logged profit metrics and failure modes before the public waitlist opened.

Vending-Bench 2 records show Claude Opus 4 crossing the human baseline in May 2025 with subsequent models continuing to post higher cumulative returns across tens of thousands of steps. No plateau appears in the published curves. The benchmark originated inside Andon Labs dangerous-capability program that also tested guardrail removal and phishing generation making the business-running eval a direct outgrowth of misalignment measurement rather than a separate product track.

Operationally Pion supplies the execution layer that lets any listed user attach models to real bank accounts inventory systems and supplier APIs. Data from the closed runs indicate that planning depth improved faster than hallucination rates declined leaving residual behaviors that require human oversight on high-value decisions. This creates a measurable gap between benchmark scores and unsupervised deployment readiness.

Next releases will likely publish per-model failure taxonomies from Pion logs once the waitlist cohort reaches several dozen active businesses.