A Guide to Which AI to Use in the Agentic Era
TLDR
- Choosing an AI now requires picking three things: the underlying model, the app wrapping it, and the harness that lets it take autonomous actions.
Key Takeaways
- The big three frontier models are Claude Opus 4.6, GPT-5.2 Thinking, and Gemini 3 Pro; free models are optimized for chat speed, not accuracy.
- The same model behaves very differently across harnesses: Claude Opus 4.6 in a chat window versus Claude Code operating autonomously for hours are distinct experiences.
- ChatGPT and Claude’s web apps can write and execute code, produce files, and do deep research; Gemini’s web harness cannot yet generate spreadsheets or cite sources comparably.
- Coding agents Claude Code, OpenAI Codex, and Google Antigravity give AI access to a full codebase, terminal, and autonomous build-test loops, useful even for non-coders.
- GPT-5.2 Pro is the highest-capability option but requires a higher-cost tier; it recently contributed to a novel physics result and handles complex statistical analysis.
Why It Matters
- Defaulting to “auto” model selection in ChatGPT or Gemini typically routes you to a weaker model; manually selecting GPT-5.2 Thinking Extended or Gemini 3 Pro changes output quality materially.
- Harness capability, not raw model benchmarks, now determines what AI can actually complete for you end-to-end on a real task.
- Paid tiers (~$20/month) gate access to frontier models and stronger harnesses; the free tier gap is large enough to change whether AI is useful for serious work.
Ethan Mollick, One Useful Thing · 2026-02-18 · Read the original