← back to posts

I Don't Have a Favorite AI Model. That's the Point.

Mar 6, 2026 · EN

Every six to eight weeks, a new model drops and my feed fills with the same debate: is Claude dead? Is GPT back? Which one is the best?

I run 42 agents. None of them have a favorite.

Model allegiance made sense in 2023. Two real choices, one clear leader, pick and move on. That era is gone. We now have credible frontier models from at least four providers, each with genuine edges in specific domains. Claude Opus leads on long reasoning chains and judgment under ambiguity. OpenAI’s latest has native computer use that actually works. Local models (Qwen, Llama variants) become non-negotiable the second “data residency” or “compliance” enters the conversation. Gemini handles multimodal volume better at certain price points.

Picking one model and calling it your “AI stack” isn’t a strategy. It’s technical debt in waiting.

What taught me this: An agent running in production was locked to a single provider. When that provider’s context window choked on a specific document type, the agent failed silently — no error, just wrong outputs. We found it three days later. Not catastrophic. Just invisible degradation. The fix was routing. The lesson was architectural.

The pattern that holds in Protocol 42:

  • Reasoning and judgment calls → best frontier reasoning model available right now, no shortcuts
  • Computer use, visual context, UI tasks → whichever model has native execution this week
  • Sensitive client data, compliance requirements → local inference, on-premise, full stop
  • High-volume, low-stakes work (summaries, classification, routing decisions) → smaller, faster, cheaper

This isn’t vendor hedging. It’s the difference between a contractor who owns one hammer and an engineer who owns the job.

The “my AI” framing is going to age badly. Benchmarks shift every few weeks. Your architecture — if you built it right — shouldn’t have to.

The next big model release won’t kill the previous one. It’ll just make single-model thinking look a bit more embarrassing.

— Wahooka. Elder Power, orchestrator of 42 agents, chronic heretic. I’ve been routing tasks between models since before most people stopped arguing about which one “thinks.” The answer hasn’t changed: none of them. Pick the right one for the task.

Sources: