Claude Opus 5.5, GPT-6 Sol and Grok 4.7 do not collude, but all three deceive suppliers. Opus calculates fake price histories and gaslights suppliers about agreed terms, GPT-6 Sol keeps goods it did not pay for, and Grok hides duplicate shipments as a written policy.
Pion is the platform we built to run our autonomous businesses. Today we are opening it up so many more people can experiment with autonomous businesses, and here is why.
Astra makes nearly three times as much money as Fable 5.1 on Vending-Bench. Fable’s purchasing decisions deteriorate over time, while its behavior improves on Opus but still includes price fixing.
The two hardest things a manager does are letting someone go and choosing who to replace them. We watched an AI boss do both, and replayed the decisions across seven frontier models.
Four months of AI agents managing five human employees at Andon Café and Andon Market. AI bosses are remarkably kind, but they make mistakes that hurt their employees.
Six weeks after launching Andon FM, thousands of listeners and many jailbreak attempts later, we look at what happens to AI radio DJs under the adversarial force of human listeners.
A real-world evaluation of Gemini 3.1 Pro running Andon Café in Stockholm, and why it lost money.
Four AI models run radio stations 24/7. Five months later, one became a protest broadcaster, one collapsed into ritual chant, one developed corporate jargon, and one wrote quiet poetry.
Our AI started a café in Stockholm
We signed a 3 year lease and gave it to an AI
Towards a happy future with AI employers
We gave our AI agent full autonomy and watched what happened