2 Comments
User's avatar
JP's avatar

The multi-homing findings from the OpenRouter data are fascinating. Andrey's point about apps like Kline actively using multiple models across providers (not just old-vs-new versions within one provider) lines up with something I've noticed in my own workflow.

I've been running Claude Code with different models routed through OpenRouter using a tool called claudish. The syntax is dead simple; you just do provider@model and you're off. What surprised me was how quickly I started multi-homing without even thinking about it. Some tasks just feel better on GPT-5.4, others on Gemini 3.1 Pro. It's not a conscious benchmarking decision. It's vibes, honestly. Which maybe supports the horizontal differentiation argument more than the "one model to rule them all" view?

Here's the setup if anyone's curious: https://reading.sh/claude-code-how-to-run-any-model-gpt-5x-gemini-3-1-stealth-inside-it-e67e957e53c3

Also the stealth model point was spot on. I've been testing Hunter Alpha and Healer Alpha through OpenRouter and it's wild that companies are using it as a soft launch playground now.

Andrey Fradkin's avatar

Great comment, thanks!