Model-agnostic vs vertically integrated: AI usage costs in practice

Really interesting looking at AI usage costs in my day-to-day agentic work.

I’m watching two experiments closely: model-agnostic versus vertically integrated.

My model-agnostic experiment is in Cursor. I use it for big chunks of my daily work, from commercial strategy and architecture right down to coding. I can choose from the leading frontier models: Anthropic, OpenAI, Google, and others. Or I can use Auto mode and let Cursor decide.

Which raises the question: does Auto choose the model that delivers the best value result for me, or the one that best serves Cursor’s gross margin? Or some combination of both? 🤷‍♂️

My vertically integrated experiment is Anthropic’s Claude. Claude only runs Anthropic models, so there’s no third-party model provider taking a margin cut. Sounds promising, unless that tight vertical integration just extracts more margin from me as a customer?

Again, dunno. 🤔

A proper side-by-side comparison across a benchmark set of tasks would settle this. If anyone has done rigorous work here, I’d be interested to see it.

In the meantime, I’ll continue as a one-man AI guinea pig.