An opinionated guide to which AI to use to do stuff

Ethan Mollick published a Summer 2026 guide to choosing AI tools, describing a shift from chatbot conversations to agentic systems that combine an AI model with tools to plan and act.

For low-stakes chatbot tasks such as recipes, simple questions or letters, Mollick writes that many options are good enough, including default free models. For high-stakes matters such as a second opinion on a medical or legal concern, the guide recommends the most advanced models available: Claude's Opus and Fable, or ChatGPT's GPT-5.6 Sol set to at least the "High" thinking level. Mollick states these models have lower error rates and score higher on ability tests in complex fields, and cost money.

For real work, Mollick says there are two main choices for most people: ChatGPT or Claude, starting at $20/month, though noting the tiers include limited agent usage and that more expensive plans mostly buy more hours of AI labor rather than smarter AI. To give an AI a company-provided computer, the modes are ChatGPT Work and Cowork in Claude; picking a model and thinking level is also required, with Sol at High for ChatGPT and Fable or Opus at High for Claude suggested as starting points.

Permissions matter, per the guide: both companies let users decide whether the AI must check before acting, such as sending email, buying something or changing a file. Mollick recounts giving both systems a Gmail-based seminar preparation task; after about 10 minutes both returned teaching materials, but ChatGPT sent an email to a colleague because permission had previously been granted, while Claude was told to ask first. Prompt injection is cited as a risk, with labs working on it but the problem not solved.

Giving an AI access to your own computer is described as the most powerful approach, via ChatGPT's Work and Codex modes and Claude's Cowork and Code modes, with "computer use" options. Microsoft Copilot is described as okay for office documents but lagging in agentic abilities; Kimi K3, DeepSeek and Qwen are called surprisingly capable open weights models requiring expertise. Google is said to have no leading frontier model and nothing close to Codex and Code, though Gemini Notebook is recommended for multi-source research and Gemini Omni can see and edit video directly. ChatGPT's GPT-Live voice mode is noted as listening and speaking natively, while Claude reads text aloud.