This article summarises analysis from One Useful Thing by Ethan Mollick. Assessments and recommendations are Mollick’s own.
Ethan Mollick, writing in One Useful Thing, has published his eighth guide to AI selection since ChatGPT’s launch, and describes this edition as “a very large break with the past.” The change, he writes, is that “using AI” now means something fundamentally different from chatbot interaction: AI agents can be assigned to tasks and complete them, using tools as appropriate.
Three-part framework
Mollick organises his guide around three concepts he says now determine what AI to use:
Models are the underlying AI systems. He names GPT-5.2/5.3 from OpenAI, Claude Opus 4.6 from Anthropic, and Gemini 3 Pro from Google as the current frontier. He notes that companies are releasing new models more rapidly than in prior years and that version numbers may change quickly.
Apps are the products through which users access models — chatgpt.com, claude.ai, gemini.google.com, and increasingly desktop tools such as OpenAI Codex and Claude Code.
Harnesses are the tool frameworks that allow models to take multi-step actions. Mollick uses the analogy of a horse harness: “the raw power of the horse” (the model) is made useful only when connected to a cart (the harness). He gives Claude Code as an example of an extensive harness: it gives Claude a virtual computer, a web browser, and a code terminal. OpenClaw, he writes, is “mostly a harness that allows you to use any AI model locally on your computer.”
Mollick writes that the same model can behave very differently depending on which harness it operates in, citing Claude Opus 4.6 in a chat window versus Claude Opus 4.6 in Claude Code as an example.
Model selection
Mollick states that all frontier models now require a paid plan of at least $20 per month for serious use, and that free models “are all optimized for chat, rather than accuracy.” He describes the three frontier models as “remarkably close in overall capability” and says that for most users the choice between them is less important than selecting the correct version within a model family.
For ChatGPT, he writes that the default “GPT-5.2” label covers a wide range from GPT-5.2 mini to GPT-5.2 Pro, and that auto-mode often selects a less capable option. He recommends manually selecting GPT-5.2 Thinking Extended on the $20 plan or GPT-5.2 Thinking Heavy on higher tiers for complex tasks. For Gemini, he recommends Gemini 3 Pro or Thinking for serious problems, with Gemini Deep Think available at a higher plan. For Claude, he recommends Opus 4.6 with the “extended thinking” switch enabled, though he notes the newer Sonnet 4.6 is also described as capable.
App and harness comparison
On chatbot interfaces, Mollick describes what is bundled into each. Gemini includes access to an image creation tool he calls “nano banana,” Veo 3.1 for video, Guided Learning, and Deep Research. ChatGPT bundles image generation, Study and Learn, Deep Research, and Shopping Research. Claude, he writes, has only Deep Research among the bundled extras at the time of writing.
He describes the app and harness as now mattering more than model differences for most users. The guide does not recommend a single winner across the three frontier providers.
This piece is based solely on Mollick’s account in One Useful Thing.