Big Three Battle: GPT-5.3 vs Claude Opus 4.6 vs Gemini 3 Pro Heats Up!
The same Claude Opus 4.6 model delivers vastly different results depending on the app and harness it uses.
AI researcher Ethan Mollick's new guide marks a fundamental shift in how to evaluate and use AI. The era of simple chatbot conversations is over; we've entered the 'agentic era' where AIs can autonomously complete tasks using tools. Mollick argues you must now assess three distinct layers: the Model (the AI's core intelligence, like GPT-5.3 or Claude Opus 4.6), the App (the product interface, like chatgpt.com or Claude Code), and the Harness (the system that grants the model tool-use and autonomy, like the one in Claude Code). He demonstrates that the exact same Claude Opus 4.6 model produces outdated info in a basic chat, provides sourced answers on claude.ai, and delivers sophisticated analysis in Claude Cowork. This tripartite framework is critical for professionals to deploy AI agents effectively for complex work.
- The AI landscape has shifted from chatbots to 'agents' that can autonomously use tools and complete multi-step tasks.
- Performance now depends on three layers: Model (intelligence), App (interface), and Harness (tool-use system), with the same model behaving differently in each.
- Key examples include Claude Opus 4.6 delivering basic chat, web search on claude.ai, or autonomous coding in Claude Code, based on the harness.
Why It Matters
Professionals must now evaluate the harness and app, not just the model, to build effective AI agents for real work.