ChatGPT Work Exposes OpenAI’s Agentic Bet

OpenAI is turning ChatGPT from a conversational surface into an execution platform; the real differentiator is not another model name, but the permissions model around getting work done.

• –Cloud and local modes create materially different security boundaries: internet access and hosted tools versus local files, apps, and programs.
• –The feature set approaches Willison’s “lethal trifecta” of private data, untrusted content, and external communication, making least privilege and approval gates essential.
• –Sol, Terra, Luna, and multiple reasoning levels offer cost-performance flexibility but risk overwhelming users with configuration choices.
• –Codex’s underlying agent technology now reaches beyond developers, potentially making one execution stack serve both technical and knowledge-work workflows.
• –Developers should evaluate connectors, persistence, browser access, scheduled tasks, and auditability—not just model quality.