Model loyalty is a trap
After hundreds of thousands of tokens this year, there's no one model to rule them all.
Hot take after burning through hundreds of thousands of tokens this year: model loyalty is a trap.
There's no one model to rule them all. No single agent. Switching is the skill now, not a sign you haven't figured it out yet.
The actual edge: knowing what's new and knowing what already works. Routing through OpenRouter, tracking real cost and latency through Helicone, picking the cheapest model that still passes evals on a given task. Switching like other people switch tabs.
If your stack assumes one provider, you're paying a tax, in dollars, in lock-in, and in being late to the next model that's better at your specific task.
Originally posted on LinkedIn.
Next note: Prompting is the new core skill
