Ethan Mollick's practical guide to which AI to actually use

Ethan Mollick's latest is a straight answer to the question people keep asking him: which AI should I use? His framing is by stakes, not by brand. For low-stakes tasks like casual questions or simple drafting, he says almost any free model is now good enough and the choice barely matters. For high-stakes decisions, meaning medical, legal, or hard professional work, he tells you to pay for the strongest models, Claude's Opus and Fable or OpenAI's GPT-5.6 Sol, and turn the thinking level up, because the error rates are lower where it counts.

For real work, he narrows the field to two: ChatGPT or Claude, both starting at $20 a month. What you get for that, he argues, is not a better chatbot but agents that can use a computer for you. To make the point concrete, he had GPT-5.6 Sol check all 195 references in his forthcoming book, and it finished in about thirty minutes. He places Google's Gemini as strong for research across many sources through its Notebook feature but still behind on frontier reasoning, Microsoft Copilot as lagging on agent abilities, and open-weight models as capable but only if you have the expertise to run them.

The through-line is a nudge more than a ranking: pay the $20, hand over work you would normally do yourself, then check the output carefully. The full guide is here.

Why it matters

If you are still using a free model for work that matters, Mollick's case is that you are leaving the useful part on the table: a $20 plan buys an agent that can do multi-hour tasks, and the skill to learn now is delegating and verifying, not prompting.