AI guide shifts from chat to agent systems — Gemini falls off list

Ethan Mollick has been publishing and updating a practical guide to which AI tools to use for different tasks, and the evolution of his recommendations over the past year tells a clear story about how AI capabilities have shifted from interesting but limited to truly transformative.

A year ago, Mollick's recommendations revolved around conversational chat — ChatGPT, Claude, and Gemini as the main players, with specialized models including o3, Claude 4 Opus, and Gemini 2.5 Pro, plus Deep Research as a useful alternative mode. The decision matrix was straightforward: which model answered your question best?

Today, the emphasis has swung sharply toward agentic systems — AI that is capable of doing the equivalent of many hours of real human work in one go. This is a fundamental shift in what users are trying to optimize for. Rather than asking the AI a question, users are now handing the AI a goal and letting it execute autonomously. Gemini has dropped off Mollick's list entirely, because Google has not yet built an established product in what he calls the "Codex/ChatGPT Work/Cowork category" — the agent space. Gemini Spark, which might have served that role, "has yet to prove itself," according to Mollick.

The most powerful way to use AI, Mollick explains, is to give it access to your computer. To do this, you download the ChatGPT or Claude apps and select an agent mode. ChatGPT offers Work and Codex modes; Claude offers Cowork and Code modes. Mollick acknowledges the naming is unintuitive — the names do not map onto each other in any coherent way, and some of the names (Work and Cowork) are reused for different features in the web version of the same tools. These agent modes "have more features and capabilities because they can access your computer."

Even within ChatGPT itself, the distinction is confusing. ChatGPT Work on a mobile device differs substantially from ChatGPT Work in the desktop app, where it functions as "a less intimidating skin on top of Codex." When you use ChatGPT Work on mobile and flip it from Chat mode, the Code Interpreter container — the system that runs code — is no longer restricted from accessing the internet, which expands the range of tasks it can handle. The fact that these different surfaces exist, with overlapping names and different underlying capabilities, reflects the pace at which agentic features are being rolled out: the infrastructure is outpacing the clarity of the interface.

A year ago, Ethan Mollick's recommendations centered on conversational AI models — ChatGPT, Claude, Gemini — with specialized models like o3, Claude 4 Opus, and Gemini 2.5 Pro. The landscape has shifted dramatically toward agentic systems, where the distinction between different AI tools now hinges not on model power alone but on their ability to autonomously take actions on behalf of the user. This reflects a maturation in how people actually use AI: rather than asking questions and copying answers, users are delegating multi-step work directly to the system.

The removal of Gemini signals a competitive gap at Google. While the company offers powerful foundational models, it has not yet established a product category that competes directly with ChatGPT's Work/Codex or Claude's Cowork/Code offerings. Gemini Spark, which might have filled that role, remains unproven. For users seeking an agent, OpenAI and Anthropic have moved faster to productize the capability.

The confusing naming scheme across platforms — Work, Cowork, Codex, Code — and the non-obvious difference between ChatGPT's mobile Work mode and desktop Codex reflects the speed at which these features are being rolled out. The fact that mobile Work removes internet restrictions from the Code Interpreter, while desktop Work is described as a skin on Codex, suggests the underlying infrastructure is maturing faster than the UX and naming conventions that surface it to users.