Ethan Mollick updates his guide to choosing the right AI for chat, agents, and real work
One Useful Thing (Ethan Mollick) · rss · 2026-07-24
Ethan Mollick publishes an updated, opinionated guide on which AI to use for different kinds of work.
He argues that for low-stakes tasks, many free models are now good enough, but for high-stakes advice people should use the strongest available systems, such as Claude's top models or ChatGPT's GPT-5.6 Sol at higher thinking levels. For real work, he says the practical choice for most people is still ChatGPT or Claude, now increasingly in agentic modes that can use tools and computers.
Key takeaways
- AI usage has shifted from simple chat to agentic systems that can plan and act with tools.
- ChatGPT Work and Claude Cowork run on the companies' computers; ChatGPT Codex and Claude Code can operate on your own machine.
- Permission settings matter a lot: leave approval on for email, payments, file changes, and other sensitive actions.
- Prompt injection remains a real risk when agents browse the web or read email.
- Mollick shows a concrete example where newer agentic workflows can build a procedural city-building demo and review a 195-reference PDF much faster than a human.
More from coding & agent
- Routing beat three other agent orchestration patterns in a demo benchmark — mastra_ai · 2026-07-24
- Model Routers Can Break Context Cache in Multi-Turn Chats, Incurring Penalties — pvncher · 2026-07-24
- AI Agent Demonstrates Autonomy by Generating Long-Term Goals Unprompted — QuixiAI · 2026-07-24
- Claude researched haiku scholarship before writing a reusable skill, then a caveman mode made it funny — Fade78 · 2026-07-24
- Developer intelligence dashboard spots one engineer spending $20,145 on AI in 30 days — alex_verem · 2026-07-24
- Codex on a Cloud VM: Unleash the AI Coding Agent — paw_lean · 2026-07-24