An opinionated guide to which AI to use to do stuff
| Source: One Useful Thing (Ethan Mollick)
Tags: Claude, ChatGPT, GPT-5.6, Opus, Fable, agentic AI, Ethan Mollick, model selection
Wharton's Ethan Mollick updates his practitioner AI guide for Summer 2026, naming Claude (Opus/Fable) and GPT-5.6 Sol as the only two serious choices for real work — and declaring that agentic systems, not chatbots, are now the baseline for anyone trying to get things done.
Details
Ethan Mollick, whose 'One Useful Thing' Substack is widely read by practitioners and researchers, has published his Summer 2026 AI guide. The core argument: the fundamental shift since his last edition is the move from chatbot to agentic AI — systems that can do 'many hours of real human work in one go' by combining model intelligence with tool access and a computer to operate. For low-stakes tasks, Mollick says any free model is fine. For high-stakes use cases — medical second opinions, legal analysis — he recommends only Claude's top models (Opus and Fable) or ChatGPT's GPT-5.6 Sol at 'High' thinking levels, citing meaningfully lower error rates and stronger performance on complex professional domains. Both platforms start at 0/month. To illustrate the pace of change, Mollick compares a brutalist city building game he coded with a prompt via GPT-5 one year ago to the same prompt run through GPT-5.6 Sol in Codex today — describing the difference as 'quite stark.' The piece also briefly covers Google but positions it as a secondary option for most use cases.