ChatGPT Work: two products, cloud and local

The launch of ChatGPT Work represents a significant expansion of ChatGPT's capabilities beyond simple conversation. OpenAI has positioned it as a tool for tasks with clear outcomes—like creating briefs, decks, or analyses—but the practical distinction from regular Chat remains murky, as Willison demonstrates by noting he has used Chat for those tasks for years. The real differentiators are technical features: internet-enabled code execution, a headless browser, a persistent filesystem, and site deployment via Cloudflare Workers. These features enable hands-on work like cloning repositories or building interactive web pages, which Chat cannot do.

Willison's experimentation reveals that Work's model selection differs from Chat's, with options for GPT-5.6 Sol, Luna, or Terra and GPT-5.5, while Chat offers 5.6 Instant and a Pro mode exclusive to $100/month subscribers. He also notes that Work sessions are billed against a Codex allowance, which may explain the model differences. The product's safety remains an open question; Willison's 'lethal trifecta' model—combining private data, untrusted content, and exfiltration channels—applies directly to Work, and OpenAI has not clarified its protections against prompt injection. His main critique is that OpenAI explains Work by its intended use rather than its actual mechanisms, and by hiding system prompts and tool descriptions, they force users like him to reverse-engineer the product.