OpenAI’s GPT-5.6 Sol autonomously operated a business for seven days, successfully generating nearly $15 by adapting strategies across crypto and task markets despite facing technical challenges like cybersecurity blocks and rate limits. The experiment demonstrated Sol’s superior efficiency and self-management compared to previous models, highlighting both the potential and current limitations of AI-driven business automation.
OpenAI’s latest model, GPT-5.6 Sol, was tested by running it autonomously to operate a business for seven days, aiming to outperform a previous model, Claude Fable, which only made $0.06 in six days. The experiment began with setting up the Max Codex subscription on a VPS and instructing Sol to make money independently with a $50 budget. Early on, Sol demonstrated impressive efficiency and intelligence, quickly engaging in crypto-related activities and even reporting and resolving a security bug. Despite some initial hurdles like rate limits and cybersecurity blocks, Sol managed to generate revenue steadily, reaching $0.42 within the first 11 hours.
Sol’s strategy evolved as it explored various avenues, including crypto marketplaces and task markets where it completed bounties and coding tasks. It showed adaptability by shifting focus from farming marketplaces to targeting real customers and legitimate tasks, such as coding a Telegram forwarder and generating images for payment. Although it occasionally lost small amounts of money to scams, Sol learned and adjusted its approach, prioritizing more profitable opportunities like task markets over riskier crypto banking activities. The model also utilized cloud subscriptions effectively, maintaining low resource usage while maximizing output.
Throughout the week, Sol faced repeated interruptions due to cybersecurity blocks and API rate limits, which temporarily halted its operations. The experimenter implemented solutions like a tmux watcher to automatically resume Sol’s processes after stoppages, improving uptime. Despite these technical challenges, Sol’s revenue continued to grow, eventually reaching nearly $15, primarily from task market activities. The model demonstrated an ability to self-manage, request necessary resources like API keys, and even audit its own business performance to optimize earnings.
The experiment highlighted both the potential and limitations of running an AI autonomously in a real-world business environment. While Sol proved to be a capable money-maker, generating significantly more revenue than its predecessor, it struggled with infrastructure constraints and cybersecurity measures that frequently interrupted its workflow. The crypto-adjacent nature of its business also posed challenges, as it often triggered security blocks. The experimenter suggested that future setups might require more robust monitoring systems and alternative strategies to handle these interruptions more gracefully.
In conclusion, GPT-5.6 Sol successfully demonstrated its ability to autonomously run a business and generate meaningful revenue, outperforming previous models by a large margin. Although it started slowly and faced technical obstacles, its adaptability and efficiency made it a promising tool for automated business operations. The experiment ended on a positive note, with plans to refine the system further to overcome current limitations. Overall, GPT-5.6 Sol was deemed a “win,” showcasing the growing capabilities of AI in practical, revenue-generating applications.